Initial backend deployment

Dockerfile

@@ -0,0 +1,35 @@
+FROM python:3.11-slim
+
+WORKDIR /app
+
+# Install system dependencies for ffmpeg, spacy, and rendering (MUST BE ROOT)
+RUN apt-get update && apt-get install -y \
+    ffmpeg \
+    libsm6 \
+    libxext6 \
+    libgl1-mesa-glx \
+    && rm -rf /var/lib/apt/lists/*
+
+# Set up a non-root user as expected by Hugging Face Spaces
+RUN useradd -m -u 1000 user
+USER user
+ENV PATH="/home/user/.local/bin:$PATH"
+
+# Copy requirements first for Docker layer caching
+COPY --chown=user requirements.txt .
+RUN pip install --no-cache-dir -r requirements.txt
+
+# Copy everything (tribev2 package, app.py, pyproject.toml)
+COPY --chown=user . .
+
+# Install the tribev2 package with plotting extras
+RUN pip install --no-cache-dir -e ".[plotting]"
+
+# Download spacy model needed by tribev2
+RUN python -m spacy download en_core_web_sm
+
+# Expose Hugging Face default port
+EXPOSE 7860
+
+# Run the FastAPI server
+CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "7860"]

README.md

@@ -0,0 +1,37 @@
+---
+title: TRIBE v2 NeuroMarketer API
+emoji: 🧠
+colorFrom: indigo
+colorTo: green
+sdk: docker
+pinned: false
+---
+
+# TRIBE v2 — Cloud Inference API
+
+This Hugging Face Space runs the TRIBE v2 (facebook/tribev2) meta-model as a FastAPI server, predicting brain-level neural responses to text and video content.
+
+## API Endpoints
+
+| Method | Endpoint | Description |
+|--------|----------|-------------|
+| GET | `/` | Health check |
+| POST | `/api/analyze-text` | Analyze text content (form field: `text`) |
+| POST | `/api/analyze-video` | Analyze video content (form field: `video`, MP4 upload) |
+
+## Response Format
+
+```json
+{
+  "success": true,
+  "htmlData": "<div>...brain heatmap as base64 image...</div>",
+  "statusInfo": "Analyzed 12 timesteps successfully."
+}
+```
+
+## Environment Variables
+
+| Variable | Description | Default |
+|----------|-------------|---------|
+| `ALLOWED_ORIGINS` | Comma-separated CORS origins | `*` |
+| `PORT` | Server port | `7860` |

app.py

@@ -0,0 +1,138 @@
+import os
+import sys
+import base64
+import tempfile
+from pathlib import Path
+from io import BytesIO
+
+from fastapi import FastAPI, Form, File, UploadFile
+from fastapi.middleware.cors import CORSMiddleware
+from pydantic import BaseModel
+import static_ffmpeg
+static_ffmpeg.add_paths()
+
+import matplotlib
+matplotlib.use("Agg")  # Non-interactive backend for server
+import matplotlib.pyplot as plt
+
+# Import TRIBE v2
+from tribev2.demo_utils import TribeModel
+from tribev2.plotting import PlotBrain
+
+# Initialize app and model
+app = FastAPI(title="TRIBE v2 Cloud API")
+
+# CORS: Allow Vercel frontend and local dev
+ALLOWED_ORIGINS = os.environ.get("ALLOWED_ORIGINS", "*").split(",")
+
+app.add_middleware(
+    CORSMiddleware,
+    allow_origins=ALLOWED_ORIGINS,
+    allow_credentials=True,
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
+print("Initializing TRIBE v2 Model on Hugging Face Space...")
+model = TribeModel.from_pretrained("facebook/tribev2")
+print("Model initialized successfully!")
+
+plotter = PlotBrain(mesh="fsaverage5")
+
+
+@app.get("/")
+async def health_check():
+    """Health check endpoint for HF Space monitoring."""
+    return {"status": "ok", "model": "TRIBE v2", "message": "API is running."}
+
+
+@app.post("/api/analyze-text")
+async def analyze_text(text: str = Form(...)):
+    print(f"Received text analysis request: {len(text)} characters")
+
+    with tempfile.NamedTemporaryFile(suffix=".txt", mode="w", delete=False) as f:
+        f.write(text)
+        temp_path = f.name
+
+    try:
+        events = model.get_events_dataframe(text_path=temp_path)
+        preds, segments = model.predict(events, verbose=False)
+
+        # Plot using matplotlib (returns a Figure)
+        fig = plotter.plot_timesteps(
+            preds,
+            segments=segments,
+            plot_every_k_timesteps=1,
+            views=["left", "right"],
+            norm_percentile=95,
+        )
+
+        # Save to base64
+        buf = BytesIO()
+        fig.savefig(buf, format="png", bbox_inches="tight", dpi=150)
+        buf.seek(0)
+        img_base64 = base64.b64encode(buf.read()).decode("utf-8")
+        plt.close(fig)
+
+        html_data = f'<div style="text-align: center; width: 100%; height: 100%; overflow: auto; background: #000; padding: 20px;"><img src="data:image/png;base64,{img_base64}" style="max-width: 100%; height: auto; border-radius: 8px;" /></div>'
+
+        return {
+            "success": True,
+            "htmlData": html_data,
+            "statusInfo": f"Analyzed {len(preds)} timesteps successfully.",
+        }
+
+    except Exception as e:
+        print("Error:", e)
+        return {"success": False, "error": str(e)}
+    finally:
+        os.remove(temp_path)
+
+
+@app.post("/api/analyze-video")
+async def analyze_video(video: UploadFile = File(...)):
+    print(f"Received video analysis request: {video.filename}")
+
+    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as f:
+        content = await video.read()
+        f.write(content)
+        temp_path = f.name
+
+    try:
+        events = model.get_events_dataframe(video_path=temp_path)
+        preds, segments = model.predict(events, verbose=False)
+
+        fig = plotter.plot_timesteps(
+            preds,
+            segments=segments,
+            plot_every_k_timesteps=2,  # Plot less frequently for video to save memory
+            views=["left", "right"],
+            norm_percentile=95,
+        )
+
+        buf = BytesIO()
+        fig.savefig(buf, format="png", bbox_inches="tight", dpi=150)
+        buf.seek(0)
+        img_base64 = base64.b64encode(buf.read()).decode("utf-8")
+        plt.close(fig)
+
+        html_data = f'<div style="text-align: center; width: 100%; height: 100%; overflow: auto; background: #000; padding: 20px;"><img src="data:image/png;base64,{img_base64}" style="max-width: 100%; height: auto; border-radius: 8px;" /></div>'
+
+        return {
+            "success": True,
+            "htmlData": html_data,
+            "statusInfo": f"Analyzed {len(preds)} timesteps of video successfully.",
+        }
+
+    except Exception as e:
+        print("Error:", e)
+        return {"success": False, "error": str(e)}
+    finally:
+        os.remove(temp_path)
+
+
+if __name__ == "__main__":
+    import uvicorn
+
+    port = int(os.environ.get("PORT", 7860))
+    uvicorn.run("app:app", host="0.0.0.0", port=port, reload=False)

pyproject.toml

@@ -0,0 +1,67 @@
+[build-system]
+requires = ["setuptools>=61.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "tribev2"
+version = "0.1.0"
+description = "Deep multimodal brain encoding"
+readme = "README.md"
+requires-python = ">=3.11"
+license = {file = "LICENSE"}
+authors = [{name = "Meta Platforms, Inc."}]
+
+dependencies = [
+  "neuralset==0.0.2",
+  "neuraltrain==0.0.2",
+  "torch>=2.5.1,<2.7",
+  "numpy==2.2.6",
+  "torchvision>=0.20,<0.22",
+  "x_transformers==1.27.20",
+  "einops",
+  "pyyaml",
+  "moviepy>=2.2.1",
+  "huggingface_hub",
+  "gtts",
+  "langdetect",
+  "spacy",
+  "soundfile",
+  "pip",
+  "Levenshtein",
+  "julius",
+  "transformers"
+]
+
+[project.urls]
+Homepage = "https://github.com/facebookresearch/tribev2"
+Repository = "https://github.com/facebookresearch/tribev2"
+
+[project.optional-dependencies]
+  plotting = [
+    "nibabel",
+    "matplotlib",
+    "seaborn",
+    "colorcet",
+    "nilearn",
+    "scipy",
+    "pyvista",
+    "scikit-image",
+  ]
+  training = [
+    "nibabel",
+    "torchmetrics",
+    "wandb",
+    "lightning",  
+  ]
+  test = [
+    "pytest",
+  ]
+
+[tool.black]
+line-length = 88
+
+[tool.isort]
+profile = "black"
+
+[tool.setuptools.packages.find]
+include = ["tribe*"]

requirements.txt

@@ -0,0 +1,5 @@
+fastapi==0.104.1
+uvicorn==0.24.0
+pydantic==2.5.2
+python-multipart==0.0.6
+static-ffmpeg==2.5

tribev2/__init__.py

@@ -0,0 +1,9 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from tribev2.demo_utils import TribeModel
+
+__all__ = ["TribeModel"]

tribev2/demo_utils.py

@@ -0,0 +1,392 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""TribeModel for inference and utilities for building event DataFrames."""
+
+import logging
+import typing as tp
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+import pydantic
+import requests
+import torch
+import yaml
+from einops import rearrange
+from exca import ConfDict, TaskInfra
+from tqdm import tqdm
+
+logger = logging.getLogger(__name__)
+logger.setLevel(logging.INFO)
+if not logger.handlers:
+    _handler = logging.StreamHandler()
+    _handler.setFormatter(logging.Formatter("%(levelname)s - %(message)s"))
+    logger.addHandler(_handler)
+from neuralset.events.transforms import (
+    AddContextToWords,
+    AddSentenceToWords,
+    AddText,
+    ChunkEvents,
+    ExtractAudioFromVideo,
+    RemoveMissing,
+)
+from neuralset.events.utils import standardize_events
+
+from tribev2.eventstransforms import ExtractWordsFromAudio
+from tribev2.main import TribeExperiment
+
+VALID_SUFFIXES: dict[str, set[str]] = {
+    "text_path": {".txt"},
+    "audio_path": {".wav", ".mp3", ".flac", ".ogg"},
+    "video_path": {".mp4", ".avi", ".mkv", ".mov", ".webm"},
+}
+
+
+def download_file(url: str, path: str | Path) -> Path:
+    """Download a file from *url* and save it to *path*.
+
+    Raises ``requests.HTTPError`` on non-2xx responses.
+    """
+    path = Path(path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with requests.get(url, stream=True, timeout=30) as r:
+        r.raise_for_status()
+        with open(path, "wb") as f:
+            for chunk in r.iter_content(chunk_size=128 * 1024):
+                if chunk:
+                    f.write(chunk)
+    logger.info(f"Downloaded {url} -> {path}")
+    return path
+
+
+def get_audio_and_text_events(
+    events: pd.DataFrame, audio_only: bool = False
+) -> pd.DataFrame:
+    """Run the audio/video-to-text pipeline on an events DataFrame.
+
+    Extracts audio from video, chunks long clips, transcribes words, and
+    attaches sentence/context annotations.  Set *audio_only* to ``True``
+    to skip the transcription and text stages.
+    """
+    transforms = [
+        ExtractAudioFromVideo(),
+        ChunkEvents(event_type_to_chunk="Audio", max_duration=60, min_duration=30),
+        ChunkEvents(event_type_to_chunk="Video", max_duration=60, min_duration=30),
+    ]
+    if not audio_only:
+        transforms.extend(
+            [
+                ExtractWordsFromAudio(),
+                AddText(),
+                AddSentenceToWords(max_unmatched_ratio=0.05),
+                AddContextToWords(
+                    sentence_only=False, max_context_len=1024, split_field=""
+                ),
+                RemoveMissing(),
+            ]
+        )
+    events = standardize_events(events)
+    for transform in transforms:
+        events = transform(events)
+    return standardize_events(events)
+
+
+class TextToEvents(pydantic.BaseModel):
+    """Convert raw text to an events DataFrame via text-to-speech + transcription.
+
+    The text is synthesised to audio with gTTS, then processed through
+    :func:`get_audio_and_text_events` to obtain word-level events.
+    """
+
+    text: str
+    infra: TaskInfra = TaskInfra()
+
+    def model_post_init(self, __context: tp.Any) -> None:
+        if self.infra.folder is None:
+            raise ValueError("A folder must be specified to save the audio file.")
+
+    @infra.apply()
+    def get_events(self) -> pd.DataFrame:
+        from gtts import gTTS
+        from langdetect import detect
+
+        audio_path = Path(self.infra.uid_folder(create=True)) / "audio.mp3"
+        lang = detect(self.text)
+        tts = gTTS(self.text, lang=lang)
+        tts.save(str(audio_path))
+        logger.info(f"Wrote TTS audio to {audio_path}")
+
+        audio_event = {
+            "type": "Audio",
+            "filepath": str(audio_path),
+            "start": 0,
+            "timeline": "default",
+            "subject": "default",
+        }
+        return get_audio_and_text_events(pd.DataFrame([audio_event]))
+
+
+class TribeModel(TribeExperiment):
+    """High-level inference wrapper around :class:`TribeExperiment`.
+
+    Provides a simple ``from_pretrained`` / ``predict`` interface for
+    generating fMRI-like brain-activity predictions from text, audio,
+    or video inputs.
+
+    Typical usage::
+
+        model = TribeModel.from_pretrained("facebook/tribev2")
+        events = model.get_events_dataframe(video_path="clip.mp4")
+        preds, segments = model.predict(events)
+    """
+
+    cache_folder: str = "./cache"
+    remove_empty_segments: bool = True
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        checkpoint_dir: str | Path,
+        checkpoint_name: str = "best.ckpt",
+        cache_folder: str | Path = None,
+        cluster: str = None,
+        device: str = "auto",
+        config_update: dict | None = None,
+    ) -> "TribeModel":
+        """Load a trained model from a checkpoint directory or HuggingFace Hub repo.
+
+        ``checkpoint_dir`` can be either a local path containing
+        ``config.yaml`` and ``<checkpoint_name>``, or a HuggingFace Hub
+        repo id (e.g. ``"facebook/tribev2"``).
+
+        Parameters
+        ----------
+        checkpoint_dir:
+            Local directory or HuggingFace Hub repo id that contains
+            ``config.yaml`` and the checkpoint file.
+        checkpoint_name:
+            Filename of the checkpoint inside *checkpoint_dir*.
+        cache_folder:
+            Directory used to cache extracted features. Created if it
+            does not exist.  Defaults to ``"./cache"`` when ``None``.
+        cluster:
+            Cluster backend forwarded to feature-extractor infra
+            (``"auto"`` by default).
+        device:
+            Torch device string.  ``"auto"`` selects CUDA when available.
+        config_update:
+            Optional dictionary of config overrides applied after the
+            YAML config is loaded.
+
+        Returns
+        -------
+        TribeModel
+            A ready-to-use model instance with weights loaded in eval mode.
+        """
+        if cache_folder is not None:
+            Path(cache_folder).mkdir(parents=True, exist_ok=True)
+        if device == "auto":
+            device = "cuda" if torch.cuda.is_available() else "cpu"
+        checkpoint_dir = Path(checkpoint_dir)
+        if checkpoint_dir.exists():
+            config_path = checkpoint_dir / "config.yaml"
+            ckpt_path = checkpoint_dir / checkpoint_name
+        else:
+            from huggingface_hub import hf_hub_download
+
+            repo_id = str(checkpoint_dir)
+            config_path = hf_hub_download(repo_id, "config.yaml")
+            ckpt_path = hf_hub_download(repo_id, checkpoint_name)
+        with open(config_path, "r") as f:
+            config = ConfDict(yaml.load(f, Loader=yaml.UnsafeLoader))
+        for modality in ["text", "audio", "video"]:
+            config[f"data.{modality}_feature.infra.folder"] = cache_folder
+            config[f"data.{modality}_feature.infra.cluster"] = cluster
+
+        for param in [
+            "infra.workdir",
+            "data.study.infra_timelines",
+            "data.neuro.infra",
+            "data.image_feature.infra",
+        ]:
+            config.pop(param)
+        config["data.study.path"] = "."
+        config["average_subjects"] = True
+        config["checkpoint_path"] = str(config["infra.folder"]) + f"/{checkpoint_name}"
+        config["cache_folder"] = (
+            str(cache_folder) if cache_folder is not None else "./cache"
+        )
+        if config_update is not None:
+            config.update(config_update)
+        xp = cls(**config)
+
+        logger.info(f"Loading model from {ckpt_path}")
+        ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=True, mmap=True)
+        build_args = ckpt["model_build_args"]
+        state_dict = {
+            k.removeprefix("model."): v for k, v in ckpt["state_dict"].items()
+        }
+        del ckpt
+
+        model = xp.brain_model_config.build(**build_args)
+        model.load_state_dict(state_dict, strict=True, assign=True)
+        del state_dict
+        model.to(device)
+        model.eval()
+        xp._model = model
+        return xp
+
+    def get_events_dataframe(
+        self,
+        text_path: str | None = None,
+        audio_path: str | None = None,
+        video_path: str | None = None,
+    ) -> pd.DataFrame:
+        """Build an events DataFrame from exactly one input source.
+
+        Parameters
+        ----------
+        text_path:
+            Path to a ``.txt`` file. The text is converted to speech, then
+            transcribed back to produce word-level events.
+        audio_path:
+            Path to an audio file (``.wav``, ``.mp3``, ``.flac``, ``.ogg``).
+        video_path:
+            Path to a video file (``.mp4``, ``.avi``, ``.mkv``, ``.mov``,
+            ``.webm``).
+
+        Returns
+        -------
+        pd.DataFrame
+            Standardised events DataFrame with columns such as ``type``,
+            ``filepath``, ``start``, ``duration``, ``timeline``, and
+            ``subject``.
+
+        Raises
+        ------
+        ValueError
+            If zero or more than one path is provided, or if the file
+            extension does not match the expected suffixes.
+        FileNotFoundError
+            If the specified file does not exist.
+        """
+        provided = {
+            name: value
+            for name, value in [
+                ("text_path", text_path),
+                ("audio_path", audio_path),
+                ("video_path", video_path),
+            ]
+            if value is not None
+        }
+        if len(provided) != 1:
+            raise ValueError(
+                f"Exactly one of text_path, audio_path, video_path must be "
+                f"provided, got: {list(provided.keys()) or 'none'}"
+            )
+
+        name, value = next(iter(provided.items()))
+        path = Path(value)
+        suffix = path.suffix.lower()
+        if suffix not in VALID_SUFFIXES[name]:
+            raise ValueError(
+                f"{name} must end with one of {sorted(VALID_SUFFIXES[name])}, "
+                f"got '{suffix}'"
+            )
+        if not path.is_file():
+            raise FileNotFoundError(f"{name} does not exist: {path}")
+
+        if text_path is not None:
+            text = path.read_text(encoding="utf-8")
+            if not text.strip():
+                raise ValueError(f"Text file is empty: {path}")
+            return TextToEvents(
+                text=text,
+                infra={"folder": self.cache_folder, "mode": "retry"},
+            ).get_events()
+
+        event_type = "Audio" if audio_path is not None else "Video"
+        event = {
+            "type": event_type,
+            "filepath": str(path),
+            "start": 0,
+            "timeline": "default",
+            "subject": "default",
+        }
+        return get_audio_and_text_events(pd.DataFrame([event]))
+
+    def predict(
+        self, events: pd.DataFrame, verbose: bool = True
+    ) -> tuple[np.ndarray, list]:
+        """Run inference on an events DataFrame and return per-TR predictions.
+
+        Each batch is split into segments of length ``data.TR``.  When
+        ``remove_empty_segments`` is ``True`` (the default), segments that
+        contain no events are discarded.
+
+        Parameters
+        ----------
+        events:
+            Events DataFrame, typically produced by
+            :meth:`get_events_dataframe`.
+        verbose:
+            If ``True`` (default), display a ``tqdm`` progress bar.
+
+        Returns
+        -------
+        preds : np.ndarray
+            Array of shape ``(n_kept_segments, n_vertices)`` with the
+            predicted brain activity.
+        all_segments : list
+            Corresponding segment objects aligned with *preds*.
+
+        Raises
+        ------
+        RuntimeError
+            If the model has not been loaded via :meth:`from_pretrained`.
+        """
+        if self._model is None:
+            raise RuntimeError(
+                "TribeModel must be instantiated via the .from_pretrained method"
+            )
+        model = self._model
+        loader = self.data.get_loaders(events=events, split_to_build="all")["all"]
+
+        preds, all_segments = [], []
+        n_samples, n_kept = 0, 0
+        with torch.inference_mode():
+            for batch in tqdm(loader, disable=not verbose):
+                batch = batch.to(model.device)
+                batch_segments = []
+                for segment in batch.segments:
+                    for t in np.arange(0, segment.duration - 1e-2, self.data.TR):
+                        batch_segments.append(
+                            segment.copy(offset=t, duration=self.data.TR)
+                        )
+                if self.remove_empty_segments:
+                    keep = np.array([len(s.ns_events) > 0 for s in batch_segments])
+                else:
+                    keep = np.ones(len(batch_segments), dtype=bool)
+                n_kept += keep.sum()
+                n_samples += len(batch_segments)
+                batch_segments = [s for i, s in enumerate(batch_segments) if keep[i]]
+                y_pred = model(batch).detach().cpu().numpy()
+                y_pred = rearrange(y_pred, "b d t -> (b t) d")[keep]
+                preds.append(y_pred)
+                all_segments.extend(batch_segments)
+        preds = np.concatenate(preds)
+        if len(all_segments) != preds.shape[0]:
+            raise ValueError(
+                f"Number of samples: {preds.shape[0]} != {len(all_segments)}"
+            )
+        logger.info(
+            "Predicted %d / %d segments (%.1f%% kept)",
+            n_kept,
+            n_samples,
+            100.0 * n_kept / max(n_samples, 1),
+        )
+        return preds, all_segments

tribev2/eventstransforms.py

@@ -0,0 +1,272 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import contextlib
+import copy
+import logging
+import os
+import typing as tp
+import warnings
+from pathlib import Path
+
+import exca
+import neuralset.events.etypes as ev
+import pandas as pd
+import torch
+
+logger = logging.getLogger(__name__)
+from neuralset.events.transforms import EventsTransform
+from neuralset.events.transforms.utils import DeterministicSplitter
+from tqdm import tqdm
+
+SPLIT_ATTRIBUTES = {
+    "Algonauts2025Bold": "chunk",
+    "Algonauts2025": "chunk",
+    "Lebel2023Bold": "task",
+    "Nastase2020": "story",
+    "Wen2017": "seg",
+    "Wenvtwo2017": "run",
+    "Lahner2024Bold": "timeline",
+    "Vanessen2023": "run",
+    "Aliko2020": "task",
+    "Li2022": "run",
+}
+
+
+def assign_splits(
+    events: pd.DataFrame, splitter: tp.Callable[str, str]
+) -> pd.DataFrame:
+    assert events.study.nunique() == 1, "Only one study can be assigned at a time"
+    study_name = events.study.unique()[0]
+    split_by = SPLIT_ATTRIBUTES[study_name]
+    events["split_attr"] = events[split_by].astype(str)
+    values = events["split_attr"].unique()
+    # check that all rows have split attr assigned
+    unassigned_event_types = events[events.split_attr.isna()].type.unique().tolist()
+    if len(unassigned_event_types) > 0:
+        msg = f"Study {study_name}: The following events do not have a split assigned and will be removed: {unassigned_event_types}"
+        if any(
+            [
+                name.capitalize() in unassigned_event_types
+                for name in ["Fmri", "Video", "Audio", "Word"]
+            ]
+        ):
+            raise ValueError(msg)
+        else:
+            events = events[~events.type.isin(unassigned_event_types)]
+            warnings.warn(msg)
+    splits = [splitter(value) for value in values]
+    if splits and "val" not in splits:
+        splits[-1] = "val"  # need at least one val split
+    val_to_split = dict(zip(values, splits))
+    events["split"] = events["split_attr"].map(val_to_split)
+    return events
+
+
+class SplitEvents(EventsTransform):
+    val_ratio: float
+
+    def _run(self, events: pd.DataFrame) -> pd.DataFrame:
+
+        splitter = DeterministicSplitter(
+            ratios={"train": 1 - self.val_ratio, "val": self.val_ratio}, seed=42
+        )
+        tmp = []
+        for _, study_events in events.groupby("study"):
+            study_events = assign_splits(study_events, splitter)
+            tmp.append(study_events)
+        events = pd.concat(tmp)
+
+        return events
+
+
+class ExtractWordsFromAudio(EventsTransform):
+    """
+    Language is hard-coded because auto-detection in performed on first 30s of audio, which can be empty e.g. for movies.
+    """
+
+    language: str = "english"
+    overwrite: bool = False
+
+    @staticmethod
+    def _get_transcript_from_audio(wav_filename: Path, language: str) -> pd.DataFrame:
+        import json
+        import os
+        import subprocess
+        import tempfile
+
+        language_codes = dict(
+            english="en", french="fr", spanish="es", dutch="nl", chinese="zh"
+        )
+        if language not in language_codes:
+            raise ValueError(f"Language {language} not supported")
+
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+        compute_type = "float32" if device == "cpu" else "float16"
+        with tempfile.TemporaryDirectory() as output_dir:
+            logger.info("Running whisperx via uvx...")
+            cmd = [
+                "uvx",
+                "whisperx",
+                str(wav_filename),
+                "--model",
+                "large-v3",
+                "--language",
+                language_codes[language],
+                "--device",
+                device,
+                "--compute_type",
+                compute_type,
+                "--batch_size",
+                "16",
+                "--align_model",
+                "WAV2VEC2_ASR_LARGE_LV60K_960H" if language == "english" else "",
+                "--output_dir",
+                output_dir,
+                "--output_format",
+                "json",
+            ]
+            cmd = [c for c in cmd if c]  # remove empty args
+            env = {k: v for k, v in os.environ.items() if k != "MPLBACKEND"}
+            result = subprocess.run(cmd, capture_output=True, text=True, env=env)
+            if result.returncode != 0:
+                raise RuntimeError(f"whisperx failed:\n{result.stderr}")
+
+            json_path = Path(output_dir) / f"{wav_filename.stem}.json"
+            transcript = json.loads(json_path.read_text())
+
+        words = []
+        for i, segment in enumerate(transcript["segments"]):
+            sentence = segment["text"]
+            sentence = sentence.replace('"', "")
+            for word in segment["words"]:
+                if "start" not in word:
+                    continue
+                word_dict = {
+                    "text": word["word"].replace('"', ""),
+                    "start": word["start"],
+                    "duration": word["end"] - word["start"],
+                    "sequence_id": i,
+                    "sentence": sentence,
+                }
+                words.append(word_dict)
+
+        transcript = pd.DataFrame(words)
+        return transcript
+
+    def _run(self, events: pd.DataFrame) -> pd.DataFrame:
+        if "Word" in events.type.unique():
+            logger.warning("Words already present in the events dataframe, skipping")
+            return events
+        audio_events = events.loc[events.type == "Audio"]
+        transcripts = {}
+        for wav_filename in tqdm(
+            audio_events.filepath.unique(),
+            total=len(audio_events.filepath.unique()),
+            desc="Extracting words from audio",
+        ):
+            wav_filename = Path(wav_filename)
+            transcript_filename = wav_filename.with_suffix(".tsv")
+            if transcript_filename.exists() and not self.overwrite:
+                try:
+                    transcript = pd.read_csv(transcript_filename, sep="\t")
+                except pd.errors.EmptyDataError:
+                    transcript = pd.DataFrame()
+                    logger.warning(f"Empty transcript file {transcript_filename}")
+            else:
+                transcript = self._get_transcript_from_audio(
+                    wav_filename, self.language
+                )
+                transcript.to_csv(transcript_filename, sep="\t", index=False)
+                logger.info(f"Wrote transcript to {transcript_filename}")
+            transcripts[str(wav_filename)] = transcript
+        all_transcripts = []
+        for audio_event in audio_events.itertuples():
+            transcript = copy.deepcopy(transcripts[audio_event.filepath])
+            if len(transcript) == 0:
+                continue
+            for k, v in audio_event._asdict().items():
+                if k in (
+                    "frequency",
+                    "filepath",
+                    "type",
+                    "start",
+                    "duration",
+                    "offset",
+                ):
+                    continue
+                transcript.loc[:, k] = v
+            transcript["type"] = "Word"
+            transcript["language"] = self.language
+            transcript["start"] += audio_event.start + audio_event.offset
+            all_transcripts.append(transcript)
+
+        if all_transcripts:
+            events = pd.concat([events, pd.concat(all_transcripts)], ignore_index=True)
+        else:
+            logger.warning("No transcripts found, skipping")
+        return events
+
+
+class CreateVideosFromImages(EventsTransform):
+    fps: int = 10
+    remove_images: bool = True
+    infra: exca.MapInfra = exca.MapInfra(cluster="processpool")
+
+    @infra.apply(
+        item_uid=lambda image_event: f"{image_event.filepath}_{image_event.duration}"
+    )
+    def create_video(self, image_events: list[ev.Image]) -> tp.Iterator[ev.Video]:
+        for image_event in image_events:
+            image_filepath = Path(image_event.filepath)
+            video_filepath = (
+                Path(self.infra.uid_folder(create=True))
+                / f"{image_filepath.stem}_{image_event.duration}.mp4"
+            )
+            from moviepy import ImageClip
+
+            video_filepath.parent.mkdir(parents=True, exist_ok=True)
+            clip = ImageClip(str(image_filepath), duration=image_event.duration)
+            with (
+                open(os.devnull, "w") as devnull,
+                contextlib.redirect_stdout(devnull),
+                contextlib.redirect_stderr(devnull),
+            ):
+                clip.write_videofile(
+                    video_filepath, codec="libx264", audio=False, fps=self.fps
+                )
+            video_event = ev.Video.from_dict(
+                image_event.to_dict()
+                | {
+                    "type": "Video",
+                    "filepath": str(video_filepath),
+                    "frequency": self.fps,
+                }
+            )
+            yield video_event
+
+    def _run(self, events: pd.DataFrame) -> pd.DataFrame:
+        images = events.loc[events.type == "Image"]
+        image_events = []
+        for image in tqdm(
+            images.itertuples(), total=len(images), desc="Extracting image events"
+        ):
+            image_events.append(ev.Image.from_dict(image._asdict()))
+        video_events = [
+            video_event.to_dict() for video_event in self.create_video(image_events)
+        ]
+        events = pd.concat([events, pd.DataFrame(video_events)], ignore_index=True)
+        if self.remove_images:
+            events = events.loc[events.type != "Image"]
+        return events.reset_index(drop=True)
+
+
+class RemoveDuplicates(EventsTransform):
+    subset: str | tp.Sequence[str] = "filepath"
+
+    def _run(self, events: pd.DataFrame) -> pd.DataFrame:
+        events = events.drop_duplicates(subset=self.subset)
+        return events

tribev2/grids/configs.py

@@ -0,0 +1,60 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Named config overrides to apply on top of default_config."""
+
+import copy
+
+from exca import ConfDict
+
+from .defaults import default_config
+
+mini_config = ConfDict(copy.deepcopy(default_config))
+mini_config.update(
+    {
+        "data": {
+            "layers_to_use": None,
+            "layer_aggregation": "mean",
+            "text_feature": {
+                "model_name": "Qwen/Qwen3-0.6B",
+                "layers": 2 / 3,
+            },
+            "video_feature": {
+                "image": {
+                    "model_name": "facebook/vjepa2-vitl-fpc64-256",
+                    "layers": 2 / 3,
+                },
+            },
+            "audio_feature": {
+                "layers": 2 / 3,
+            },
+        },
+    }
+)
+
+base_config = ConfDict(copy.deepcopy(default_config))
+base_config.update(
+    {
+        "data": {
+            "text_feature": {
+                "cache_n_layers": 20,
+            },
+            "image_feature": {
+                "image": {
+                    "cache_n_layers": 20,
+                },
+            },
+            "video_feature": {
+                "image": {
+                    "cache_n_layers": 20,
+                },
+            },
+            "audio_feature": {
+                "cache_n_layers": 20,
+            },
+        },
+    }
+)

tribev2/grids/defaults.py

@@ -0,0 +1,267 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Default configuration dictionary for TRIBE v2 experiments."""
+import os
+from pathlib import Path
+
+PROJECT_NAME = "tribe_release"
+
+SLURM_PARTITION = os.getenv("SLURM_PARTITION", "")
+SLURM_CONSTRAINT = os.getenv("SLURM_CONSTRAINT", "")
+WANDB_ENTITY = os.getenv("WANDB_ENTITY", "")
+DATADIR = os.getenv("DATAPATH")
+BASEDIR = os.getenv("SAVEPATH")
+CACHEDIR = os.path.join(BASEDIR, "cache", PROJECT_NAME)
+SAVEDIR = os.path.join(BASEDIR, "results", PROJECT_NAME)
+N_CPUS = 20
+
+for path in [CACHEDIR, SAVEDIR, DATADIR]:
+    Path(path).mkdir(parents=True, exist_ok=True)
+
+text_feature = {
+    "name": "HuggingFaceText",
+    "event_types": "Word",
+    "model_name": "meta-llama/Llama-3.2-3B",
+    "aggregation": "sum",
+    "frequency": 2,
+    "contextualized": True,
+    "layers": [0, 0.2, 0.4, 0.6, 0.8, 1.0],
+    "batch_size": 4,
+}
+image_feature = {
+    "name": "HuggingFaceVideo",
+    "frequency": 2,
+    "event_types": "Video",
+    "aggregation": "sum",
+    "image": {
+        "name": "HuggingFaceImage",
+        "model_name": "facebook/dinov2-large",
+        "layers": 2 / 3,
+        "infra": {"keep_in_ram": False},
+        "batch_size": 4,
+    },
+}
+video_feature = image_feature | {
+    "clip_duration": 4,
+    "image": {
+        "name": "HuggingFaceImage",
+        "model_name": "facebook/vjepa2-vitg-fpc64-256",
+        "infra": {"keep_in_ram": False},
+        "layers": [0.75, 1.0],
+    },
+}
+audio_feature = {
+    "name": "Wav2VecBert",
+    "frequency": 2,
+    "layers": [0.75, 1.0],
+    "event_types": "Audio",
+    "aggregation": "sum",
+}
+neuro_extractor = {
+    "name": "FmriExtractor",
+    "allow_missing": True,
+    "offset": 5,
+    "frequency": 1,
+    "projection": {
+        "name": "SurfaceProjector",
+        "mesh": "fsaverage5",
+        "kind": "ball",
+        "radius": 3,
+    },
+}
+for extractor in [
+    text_feature,
+    image_feature,
+    video_feature,
+    audio_feature,
+    neuro_extractor,
+]:
+    extractor["infra"] = {
+        "cluster": "slurm",
+        "cpus_per_task": 10,
+        "folder": CACHEDIR,
+        "keep_in_ram": True,
+        "mode": "cached",
+        "min_samples_per_job": 1,
+        "max_jobs": 256,
+        "timeout_min": 60 * 12,
+        "slurm_partition": SLURM_PARTITION,
+    }
+    extractor["infra"]["version"] = "release"
+    if extractor["name"] == "FmriExtractor":
+        extractor["infra"]["max_jobs"] = 1024
+    else:
+        extractor["infra"]["gpus_per_node"] = 1
+        extractor["infra"]["slurm_constraint"] = SLURM_CONSTRAINT
+    if extractor["name"] == "HuggingFaceVideo":
+        extractor["infra"]["min_samples_per_job"] = 1
+        extractor["infra"]["max_jobs"] = 1024
+        extractor["infra"]["timeout_min"] = 60 * 24
+    if extractor["name"] == "HuggingFaceText":
+        extractor["infra"]["min_samples_per_job"] = 32
+    extractor["allow_missing"] = True
+    extractor["=replace="] = True
+
+default_config = {
+    "infra": {
+        "cluster": "slurm",
+        "slurm_partition": SLURM_PARTITION,
+        "folder": SAVEDIR,
+        "gpus_per_node": 1,
+        "cpus_per_task": N_CPUS,
+        "mem_gb": 128,
+        "timeout_min": 60 * 24 * 3,
+        "mode": "retry",
+        "slurm_constraint": SLURM_CONSTRAINT,
+        "workdir": {
+            "copied": ["neuralset", "neuraltrain", "tribev2"],
+            "includes": ["*.py", "*.txt"],
+        },
+    },
+    "data": {
+        "frequency": 2,
+        "duration_trs": 100,
+        "overlap_trs_train": 0,
+        "overlap_trs_val": 0,
+        "shuffle_val": True,
+        "num_workers": N_CPUS,
+        "layers_to_use": [0.5, 0.75, 1.0],
+        "layer_aggregation": "group_mean",
+        "study": {
+            "names": [
+                "Algonauts2025Bold",
+                "Wen2017",
+                "Lahner2024Bold",
+                "Lebel2023Bold",
+            ],
+            "path": DATADIR,
+            "query": None,
+            "infra_timelines": {
+                "folder": CACHEDIR,
+                "timeout_min": 60 * 12,
+                "min_samples_per_job": 4,
+                "max_jobs": 1024,
+                "version": "final",
+            },
+            "transforms": {
+                "extractaudio": {"name": "ExtractAudioFromVideo"},
+                "extractwords": {"name": "ExtractWordsFromAudio"},
+                "addtext": {"name": "AddText"},
+                "addsentence": {
+                    "name": "AddSentenceToWords",
+                    "max_unmatched_ratio": 0.05,
+                },
+                "addcontext": {
+                    "name": "AddContextToWords",
+                    "sentence_only": False,
+                    "max_context_len": 1024,
+                    "split_field": "",
+                },
+                "removemissing": {"name": "RemoveMissing"},
+                "chunksounds": {
+                    "name": "ChunkEvents",
+                    "event_type_to_chunk": "Audio",
+                    "max_duration": 60,
+                    "min_duration": 30,
+                },
+                "chunkvideos": {
+                    "name": "ChunkEvents",
+                    "event_type_to_chunk": "Video",
+                    "max_duration": 60,
+                    "min_duration": 30,
+                    "infra": {"backend": "Cached", "folder": CACHEDIR},
+                },
+                "query": {"name": "QueryEvents", "query": None},
+                "split": {"name": "SplitEvents", "val_ratio": 0.1},
+            },
+        },
+        "neuro": neuro_extractor,
+        "features_to_use": ["text", "audio", "video"],
+        "text_feature": text_feature,
+        "video_feature": video_feature,
+        "audio_feature": audio_feature,
+        "image_feature": image_feature,
+        "batch_size": 8,
+    },
+    "wandb_config": {
+        "log_model": False,
+        "entity": WANDB_ENTITY,
+        "project": PROJECT_NAME,
+        "group": "default",
+    },
+    "brain_model_config": {
+        "name": "FmriEncoder",
+        "low_rank_head": 2048,
+        "hidden": 1152,
+        "extractor_aggregation": "cat",
+        "layer_aggregation": "cat",
+        "combiner": None,
+        "encoder": {
+            "depth": 8,
+        },
+        "subject_layers": {"subject_dropout": 0.1},
+        "subject_embedding": False,
+        "modality_dropout": 0.3,
+    },
+    "metrics": [
+        {
+            "log_name": "pearson",
+            "name": "OnlinePearsonCorr",
+            "dim": 0,
+        },
+        {
+            "log_name": "subj_pearson",
+            "name": "GroupedMetric",
+            "metric_name": "OnlinePearsonCorr",
+            "kwargs": {"dim": 0},
+        },
+        {
+            "log_name": "retrieval_top1",
+            "name": "TopkAcc",
+            "topk": 1,
+        },
+    ],
+    "loss": {"name": "MSELoss", "kwargs": {"reduction": "none"}},
+    "optim": {
+        "name": "LightningOptimizer",
+        "optimizer": {
+            "name": "Adam",
+            "lr": 1e-4,
+            "kwargs": {
+                "weight_decay": 0.0,
+            },
+        },
+        "scheduler": {
+            "name": "OneCycleLR",
+            "kwargs": {
+                "max_lr": 1e-4,
+                "pct_start": 0.1,
+            },
+        },
+    },
+    "n_epochs": 15,
+    "limit_train_batches": None,
+    "patience": None,
+    "enable_progress_bar": True,
+    "log_every_n_steps": 5,
+    "fast_dev_run": False,
+    "seed": 33,
+}
+
+
+if __name__ == "__main__":
+    # The following can be used for local debugging/quick tests.
+
+    from ..main import TribeExperiment
+
+    exp = TribeExperiment(
+        **default_config,
+    )
+
+    exp.infra.clear_job()
+    out = exp.run()
+    print(out)

tribev2/grids/run_cortical.py

@@ -0,0 +1,44 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from exca import ConfDict
+from neuraltrain.utils import run_grid
+
+from ..main import TribeExperiment  # type: ignore
+from .configs import mini_config
+
+GRID_NAME = "cortical"
+
+update = {
+    "wandb_config.group": GRID_NAME,
+}
+
+grid = {
+    "data.study.names": [
+        "Algonauts2025Bold",
+        "Lahner2024Bold",
+        "Lebel2023Bold",
+        "Wen2017",
+        ["Algonauts2025Bold", "Lahner2024Bold", "Lebel2023Bold", "Wen2017"],
+    ],
+}
+
+
+if __name__ == "__main__":
+    updated_config = ConfDict(mini_config)
+    updated_config.update(update)
+
+    out = run_grid(
+        TribeExperiment,
+        GRID_NAME,
+        updated_config,
+        grid,
+        job_name_keys=["wandb_config.name", "infra.job_name"],
+        combinatorial=True,
+        overwrite=False,
+        dry_run=False,
+        infra_mode="force",
+    )

tribev2/grids/run_subcortical.py

@@ -0,0 +1,52 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from exca import ConfDict
+from neuraltrain.utils import run_grid
+
+from ..main import TribeExperiment  # type: ignore
+from .configs import mini_config
+
+GRID_NAME = "subcortical"
+
+update = {
+    "wandb_config.group": GRID_NAME,
+    "data.neuro": {
+        "projection": {
+            "name": "MaskProjector",
+            "mask": "subcortical",
+            "=replace=": True,
+        },
+        "fwhm": 6.0,
+    },
+}
+
+grid = {
+    "data.study.names": [
+        "Algonauts2025Bold",
+        "Lahner2024Bold",
+        "Lebel2023Bold",
+        "Wen2017",
+        ["Algonauts2025Bold", "Lahner2024Bold", "Lebel2023Bold", "Wen2017"],
+    ],
+}
+
+
+if __name__ == "__main__":
+    updated_config = ConfDict(mini_config)
+    updated_config.update(update)
+
+    out = run_grid(
+        TribeExperiment,
+        GRID_NAME,
+        updated_config,
+        grid,
+        job_name_keys=["wandb_config.name", "infra.job_name"],
+        combinatorial=True,
+        overwrite=False,
+        dry_run=False,
+        infra_mode="force",
+    )

tribev2/grids/test_run.py

@@ -0,0 +1,47 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Quick test run on reduced data and number of epochs for CI.
+"""
+
+import os
+
+from exca import ConfDict
+
+from ..main import TribeExperiment  # type: ignore
+from .configs import mini_config
+
+update = {
+    "data.num_workers": 0,
+    "infra.cluster": None,
+    "infra.workdir": None,
+    "wandb_config": None,
+    "save_checkpoints": False,
+    "n_epochs": 3,
+    "infra.gpus_per_node": 1,
+    "infra.mode": "force",
+    "data.study.names": "Algonauts2025Bold",
+    "data.study.transforms.query.query": "subject_timeline_index<3",
+}
+
+updated_config = ConfDict(mini_config)
+updated_config.update(update)
+
+
+def test_run(config: dict) -> None:
+    task = TribeExperiment(**config)
+    task.infra.clear_job()
+    task.run()
+
+
+if __name__ == "__main__":
+    folder = os.path.join(updated_config["infra"]["folder"], "test")
+    updated_config["infra"]["folder"] = folder
+    if os.path.exists(folder):
+        import shutil
+
+        shutil.rmtree(folder)
+    test_run(updated_config)

tribev2/main.py

@@ -0,0 +1,651 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Defines the main classes used in the experiment.
+
+We suggest the following structure:
+- `Data`: configures dataset and extractors to return DataLoaders
+- `Trainer`: creates the deep learning model and exposes a `fit` and `test` methods
+- `TribeExperiment`: main class that defines the experiment to run by using `Data` and `Trainer`
+"""
+
+import gc
+import logging
+import os
+import typing as tp
+from pathlib import Path
+
+import neuralset as ns
+import numpy as np
+import pandas as pd
+import pydantic
+import torch
+import yaml
+from exca import ConfDict, TaskInfra
+from neuralset.events.etypes import EventTypesHelper
+from neuralset.events.utils import standardize_events
+from neuraltrain.losses import BaseLoss
+from neuraltrain.metrics import BaseMetric
+from neuraltrain.models import BaseModelConfig
+from neuraltrain.models.common import SubjectLayers
+from neuraltrain.optimizers.base import BaseOptimizer
+from neuraltrain.utils import BaseExperiment, WandbLoggerConfig
+from torch import nn
+from torch.utils.data import DataLoader
+
+from .eventstransforms import *  # register custom events transforms in neuralset
+from .model import *  # register custom models in neuraltrain
+from .studies import *  # register studies
+from .utils import (
+    MultiStudyLoader,
+    set_study_in_average_subject_mode,
+    split_segments_by_time,
+)
+from .utils_fmri import *  # register TribeSurfaceProjector
+
+# Configure logger
+LOGGER = logging.getLogger(__name__)
+_handler = logging.StreamHandler()
+_formatter = logging.Formatter("[%(asctime)s %(levelname)s] %(message)s", "%H:%M:%S")
+_handler.setFormatter(_formatter)
+if not LOGGER.handlers:
+    LOGGER.addHandler(_handler)
+LOGGER.setLevel(logging.INFO)
+
+
+def _free_extractor_model(extractor: ns.extractors.BaseExtractor) -> None:
+    """Delete cached GPU model from an extractor after its features are cached.
+
+    Extractors lazily load models onto GPU during ``prepare`` and keep them
+    in ``_model``.  Since results are persisted to disk, the model is no
+    longer needed afterwards and this frees VRAM for subsequent extractors.
+    """
+    targets = [extractor]
+    if hasattr(extractor, "image"):
+        targets.append(extractor.image)
+    for target in targets:
+        for attr in ("_model",):
+            obj = getattr(target, attr, None)
+            if isinstance(obj, torch.nn.Module):
+                try:
+                    delattr(target, attr)
+                except Exception:
+                    pass
+    gc.collect()
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+
+
+class Data(pydantic.BaseModel):
+    """Handles configuration and creation of DataLoaders from dataset and extractors."""
+
+    model_config = pydantic.ConfigDict(extra="forbid")
+
+    study: MultiStudyLoader
+    # features
+    neuro: ns.extractors.BaseExtractor
+    text_feature: ns.extractors.BaseExtractor | None = None
+    image_feature: ns.extractors.BaseExtractor | None = None
+    audio_feature: ns.extractors.BaseExtractor | None = None
+    video_feature: ns.extractors.BaseExtractor | None = None
+    subject_id: ns.extractors.LabelEncoder = ns.extractors.LabelEncoder(
+        event_field="subject", allow_missing=True, aggregation="first"
+    )
+    frequency: float | None = None
+    features_to_use: list[
+        tp.Literal["text", "audio", "video", "image", "context", "flow", "music"]
+    ]
+    features_to_mask: list[
+        tp.Literal["text", "audio", "video", "image", "context", "flow", "music"]
+    ] = []
+    n_layers_to_use: int | None = None
+    layers_to_use: list[float] | None = None
+    layer_aggregation: tp.Literal["group_mean", "mean"] | None = "group_mean"
+    # Dataset
+    duration_trs: int = 40
+    overlap_trs_train: int = 0
+    overlap_trs_val: int | None = None
+    batch_size: int = 64
+    num_workers: int | None = None
+    shuffle_train: bool = True
+    shuffle_val: bool = False
+    stride_drop_incomplete: bool = False
+    split_segments_by_time: bool = False
+
+    def model_post_init(self, __context):
+        super().model_post_init(__context)
+        layers_to_use = None
+        if self.n_layers_to_use is not None or self.layers_to_use is not None:
+            assert not (
+                self.n_layers_to_use is not None and self.layers_to_use is not None
+            ), "Only one of n_layers_to_use or layers_to_use can be specified"
+            if self.n_layers_to_use is not None:
+                layers_to_use = np.linspace(0, 1, self.n_layers_to_use).tolist()
+            else:
+                layers_to_use = self.layers_to_use
+        for modality in self.features_to_use:
+            extractor = getattr(self, f"{modality}_feature")
+            if hasattr(extractor, "layers"):
+                setattr(extractor, "layer_aggregation", self.layer_aggregation)
+                if layers_to_use is not None:
+                    setattr(extractor, "layers", layers_to_use)
+            if hasattr(extractor, "image") and hasattr(extractor.image, "layers"):
+                setattr(extractor.image, "layer_aggregation", self.layer_aggregation)
+                if layers_to_use is not None:
+                    setattr(extractor.image, "layers", layers_to_use)
+        if self.frequency is not None:
+            for modality in self.features_to_use:
+                extractor = getattr(self, f"{modality}_feature")
+                if hasattr(extractor, "frequency"):
+                    setattr(extractor, "frequency", self.frequency)
+
+    @property
+    def TR(self) -> float:
+        return 1 / self.neuro.frequency
+
+    def get_events(self) -> pd.DataFrame:
+        events = self.study.run()
+        events = events[events.type != "Sentence"]
+
+        cols = ["index", "subject", "timeline"]
+        event_summary = (
+            events.reset_index().groupby(["study", "split", "type"])[cols].nunique()
+        )
+        LOGGER.info("Event summary: \n%s", event_summary)
+        return events
+
+    def get_loaders(
+        self,
+        events: pd.DataFrame | None = None,
+        split_to_build: tp.Literal["train", "val", "all"] | None = None,
+    ) -> tuple[dict[str, DataLoader], int]:
+
+        if events is None:
+            events = self.get_events()
+        else:
+            events = standardize_events(events)
+
+        extractors = {}
+        for modality in self.features_to_use:
+            extractors[modality] = getattr(self, f"{modality}_feature")
+        if "Fmri" in events.type.unique():
+            extractors["fmri"] = self.neuro
+        dummy_events = []
+        for timeline_name, timeline in events.groupby("timeline"):
+            if "split" in timeline.columns:
+                splits = timeline.split.dropna().unique()
+                assert (
+                    len(splits) == 1
+                ), f"Timeline {timeline_name} has multiple splits: {splits}"
+                split = splits[0]
+            else:
+                split = "all"
+            dummy_event = {
+                "type": "CategoricalEvent",
+                "timeline": timeline_name,
+                "start": timeline.start.min(),
+                "duration": timeline.stop.max() - timeline.start.min(),
+                "split": split,
+                "subject": timeline.subject.unique()[0],
+            }
+            dummy_events.append(dummy_event)
+        events = pd.concat([events, pd.DataFrame(dummy_events)])
+        events = standardize_events(events)
+
+        extractors["subject_id"] = self.subject_id
+
+        features_to_remove = set()
+        for extractor_name, extractor in extractors.items():
+            event_types = EventTypesHelper(extractor.event_types).names
+            if not any(
+                [event_type in events.type.unique() for event_type in event_types]
+            ):
+                features_to_remove.add(extractor_name)
+        for extractor_name in features_to_remove:
+            del extractors[extractor_name]
+            LOGGER.warning(
+                "Removing extractor %s as there are no corresponding events",
+                extractor_name,
+            )
+
+        for name, extractor in extractors.items():
+            LOGGER.info("Preparing extractor: %s", name)
+            extractor.prepare(events)
+            _free_extractor_model(extractor)
+
+        # Prepare dataloaders
+        loaders = {}
+        if split_to_build is None:
+            splits = ["train", "val"]
+        else:
+            splits = [split_to_build]
+        for split in splits:
+            LOGGER.info("Building dataloader for split %s", split)
+            if split == "all" or self.split_segments_by_time:
+                split_sel = [True] * len(events)
+                shuffle = False
+                overlap_trs = self.overlap_trs_train
+            else:
+                split_sel = events.split == split
+                if split not in events.split.unique():
+                    shuffle = False
+                else:
+                    shuffle = (
+                        self.shuffle_train if split == "train" else self.shuffle_val
+                    )
+                if split == "val":
+                    overlap_trs = self.overlap_trs_val or self.overlap_trs_train
+                else:
+                    overlap_trs = self.overlap_trs_train
+
+            sel = np.array(split_sel)
+            segments = ns.segments.list_segments(
+                events[sel],
+                triggers=events[sel].type == "CategoricalEvent",
+                stride=(self.duration_trs - overlap_trs) * self.TR,
+                duration=self.duration_trs * self.TR,
+                stride_drop_incomplete=self.stride_drop_incomplete,
+            )
+            if self.split_segments_by_time:
+                LOGGER.info(f"Total number of segments: {len(segments)}")
+                segments = split_segments_by_time(
+                    segments,
+                    val_ratio=self.study.transforms["split"].val_ratio,
+                    split=split,
+                )
+                LOGGER.info(f"# {split} segments: {len(segments)}")
+            if len(segments) == 0:
+                LOGGER.warning("No events found for split %s", split)
+                continue
+            dataset = ns.dataloader.SegmentDataset(
+                extractors=extractors,
+                segments=segments,
+                remove_incomplete_segments=False,
+            )
+            dataloader = dataset.build_dataloader(
+                shuffle=shuffle,
+                num_workers=self.num_workers,
+                batch_size=self.batch_size,
+            )
+            loaders[split] = dataloader
+
+        return loaders
+
+
+class TribeExperiment(BaseExperiment):
+    """Defines the main experiment pipeline including data loading and training/evaluation."""
+
+    model_config = pydantic.ConfigDict(extra="forbid")
+
+    data: Data
+    # Reproducibility
+    seed: int | None = 33
+    # Model
+    brain_model_config: BaseModelConfig
+    # Loss
+    loss: BaseLoss
+    # Optimization
+    optim: BaseOptimizer
+    # Metrics
+    metrics: list[BaseMetric]
+    monitor: str = "val/pearson"
+    # Weights & Biases
+    wandb_config: WandbLoggerConfig | None = None
+    # Hardware
+    accelerator: str = "gpu"
+    # Optim
+    n_epochs: int | None = 10
+    max_steps: int = -1
+    patience: int | None = None
+    limit_train_batches: int | None = None
+    accumulate_grad_batches: int = 1
+    # Others
+    enable_progress_bar: bool = True
+    log_every_n_steps: int | None = None
+    fast_dev_run: bool = False
+    save_checkpoints: bool = True
+    checkpoint_filename: str = "best"
+    resize_subject_layer: bool = False
+    freeze_backbone: bool = False
+    # Eval
+    average_subjects: bool = False
+    checkpoint_path: str | None = None
+    load_checkpoint: bool = True
+    test_only: bool = False
+
+    # Internal properties
+    _trainer: tp.Any = None
+    _model: tp.Any = None
+    _logger: tp.Any = None
+
+    # Others
+    infra: TaskInfra = TaskInfra(version="1")
+
+    def model_post_init(self, __context: tp.Any) -> None:
+        super().model_post_init(__context)
+        if self.infra.folder is None:
+            msg = "infra.folder needs to be specified to save the results."
+            raise ValueError(msg)
+        # Update Trainer parameters based on infra
+        self.infra.tasks_per_node = self.infra.gpus_per_node
+        self.infra.slurm_use_srun = True if self.infra.gpus_per_node > 1 else False
+        if self.infra.gpus_per_node > 1:
+            self.metrics = [m for m in self.metrics if m.name not in ["TopkAcc"]]
+            self.data.batch_size = self.data.batch_size // self.infra.gpus_per_node
+        if self.accumulate_grad_batches > 1:
+            self.data.batch_size = self.data.batch_size // self.accumulate_grad_batches
+
+        if (
+            not (self.checkpoint_path and self.load_checkpoint)
+        ) or self.resize_subject_layer:
+            study_summary = self.data.study.study_summary()
+            self.data.subject_id.predefined_mapping = {
+                subject: i for i, subject in enumerate(study_summary.subject.unique())
+            }
+            self.brain_model_config.subject_layers.n_subjects = (
+                study_summary.subject.nunique()
+            )
+            if isinstance(self.brain_model_config.projector, SubjectLayers):
+                self.brain_model_config.projector.n_subjects = (
+                    study_summary.subject.nunique()
+                )
+
+        if self.average_subjects:
+            study_name = self.data.study.names
+            self.brain_model_config.subject_layers.average_subjects = True
+            self.brain_model_config.subject_layers.n_subjects = 0
+            if isinstance(self.brain_model_config.projector, SubjectLayers):
+                self.brain_model_config.projector.average_subjects = True
+            self.data.neuro.aggregation = "mean"
+            self.data.subject_id.predefined_mapping = None
+            if isinstance(study_name, str):
+                LOGGER.debug(f"Setting study {study_name} in average subject mode")
+                trigger_type = (
+                    "Video" if study_name in ["Wen2017", "Allen2022Bold"] else "Audio"
+                )
+                self.data.study = set_study_in_average_subject_mode(
+                    self.data.study, trigger_type=trigger_type, trigger_field="filepath"
+                )
+            else:
+                pass
+                # LOGGER.warning(
+                #     "Cannot set study in average subject mode with multiple studies"
+                # )
+
+    def _get_checkpoint_path(self) -> Path | None:
+        if self.checkpoint_path:
+            assert Path(
+                self.checkpoint_path
+            ).exists(), f"Checkpoint path {self.checkpoint_path} does not exist."
+            checkpoint_path = Path(self.checkpoint_path)
+        else:
+            checkpoint_path = Path(self.infra.folder) / "last.ckpt"
+            if not checkpoint_path.exists():
+                checkpoint_path = None
+        return checkpoint_path
+
+    def _init_module(self, model: nn.Module) -> tp.Any:
+        from .pl_module import BrainModule
+
+        checkpoint_path = self._get_checkpoint_path()
+        if (
+            self.load_checkpoint
+            and checkpoint_path is not None
+            and not self.resize_subject_layer
+        ):
+            LOGGER.info(f"Loading model from {checkpoint_path}")
+            init_fn = BrainModule.load_from_checkpoint
+            init_kwargs = {"checkpoint_path": checkpoint_path, "strict": False}
+        else:
+            init_fn = BrainModule
+            init_kwargs = {}
+
+        metrics = {
+            split + "/" + metric.log_name: metric.build()
+            for metric in self.metrics
+            for split in ["val", "test"]
+        }
+        metrics = nn.ModuleDict(metrics)
+        pl_module = init_fn(
+            model=model,
+            loss=self.loss.build(),
+            optim_config=self.optim,
+            metrics=metrics,
+            config=ConfDict(self.model_dump()),
+            **init_kwargs,
+        )
+
+        if self.resize_subject_layer:
+            LOGGER.info("Resizing subject layer")
+            checkpoint = torch.load(checkpoint_path)
+            state_dict = checkpoint["state_dict"]
+            weights = state_dict["model.predictor.weights"]
+            _, in_channels, out_channels = weights.shape
+            n_subjects = self.brain_model_config.subject_layers.n_subjects
+            if self.brain_model_config.subject_layers.subject_dropout:
+                n_subjects += 1
+            if "model.predictor.bias" in state_dict:
+                bias = state_dict["model.predictor.bias"]
+                new_bias = torch.nn.Parameter(torch.zeros(n_subjects, out_channels))
+                new_bias.data[:] = bias.mean(dim=0).repeat(n_subjects, 1)
+                state_dict["model.predictor.bias"] = new_bias
+            if self.freeze_backbone:
+                for param in pl_module.parameters():
+                    param.requires_grad = False
+            for param in pl_module.model.predictor.parameters():
+                param.requires_grad = True
+            if (
+                self.brain_model_config.low_rank_head is not None
+                and self.brain_model_config.low_rank_head != in_channels
+            ):
+                r = self.brain_model_config.low_rank_head
+                if "model.low_rank_head.weight" in state_dict:
+                    W1, W2 = (
+                        state_dict["model.low_rank_head.weight"].cpu(),
+                        state_dict["model.predictor.weights"].mean(dim=0).cpu(),
+                    )
+                    prod = torch.matmul(W1.t(), W2)
+                else:
+                    prod = state_dict["model.predictor.weights"].mean(dim=0).cpu()
+                U, S, V = torch.svd(prod)
+                U = U[:, :r]
+                S = S[:r]
+                V = V[:, :r]
+                state_dict["model.low_rank_head.weight"] = U.t()
+                state_dict["model.predictor.weights"] = torch.matmul(
+                    torch.diag(S), V.t()
+                ).repeat(n_subjects, 1, 1)
+                if "model.predictor.bias" in state_dict:
+                    state_dict["model.low_rank_head.bias"] = torch.zeros(r)
+                for param in pl_module.model.low_rank_head.parameters():
+                    param.requires_grad = True
+            else:
+                state_dict["model.predictor.weights"] = weights.mean(dim=0).repeat(
+                    n_subjects, 1, 1
+                )
+            pl_module.load_state_dict(state_dict, strict=False)
+
+        return pl_module
+
+    def _setup_trainer(
+        self, train_loader: DataLoader, override_n_devices: int | None = None
+    ) -> tp.Any:
+        import lightning.pytorch as pl
+        from lightning.pytorch.callbacks import (
+            EarlyStopping,
+            LearningRateMonitor,
+            ModelCheckpoint,
+        )
+
+        batch = next(iter(train_loader))
+        feature_dims = {}
+        for modality in self.data.features_to_use:
+            if (
+                modality in batch.data and modality not in self.data.features_to_mask
+            ):  # B, L, D, T
+                if batch.data[modality].ndim == 4:
+                    feature_dims[modality] = (
+                        batch.data[modality].shape[1],
+                        batch.data[modality].shape[2],
+                    )
+                elif batch.data[modality].ndim == 3:
+                    feature_dims[modality] = (
+                        1,
+                        batch.data[modality].shape[1],
+                    )
+                else:
+                    raise ValueError(
+                        f"Unexpected number of dimensions for modality {modality}: {batch.data[modality].ndim}"
+                    )
+            else:
+                feature_dims[modality] = None
+        if "fmri" in batch.data:  # read from fmri config
+            fmri = batch.data["fmri"]
+            n_outputs = fmri.shape[1]
+            for metric in self.metrics:
+                if hasattr(metric, "kwargs") and "num_outputs" in metric.kwargs:
+                    metric.kwargs["num_outputs"] = n_outputs
+        else:  # read from neuro config
+            if hasattr(self.data.neuro.projection, "mesh"):
+                from neuralset.extractors.neuro import FSAVERAGE_SIZES
+
+                n_outputs = 2 * FSAVERAGE_SIZES[self.data.neuro.projection.mesh]
+            else:
+                raise ValueError(
+                    f"Could not determine number of outputs for neuro extractor {self.data.neuro}"
+                )
+        brain_model = self.brain_model_config.build(
+            feature_dims=feature_dims,
+            n_outputs=n_outputs,
+            n_output_timesteps=self.data.duration_trs,
+        )
+        LOGGER.info("Extractor dims: %s", feature_dims)
+        input_data = brain_model.aggregate_features(batch)
+        LOGGER.info("Input shapes: %s", input_data.shape)
+        LOGGER.info("Target shapes: %s", n_outputs)
+        _ = brain_model(batch)
+        total_params = sum(p.numel() for p in brain_model.parameters())
+        LOGGER.info(f"Total parameters: {total_params}")
+        self._model = self._init_module(brain_model)
+        if self.monitor == "val/pearson":
+            mode = "max"
+        else:
+            mode = "min"
+        callbacks = [
+            LearningRateMonitor(logging_interval="epoch"),
+        ]
+        if self.patience is not None:
+            callbacks.append(
+                EarlyStopping(monitor=self.monitor, mode=mode, patience=self.patience)
+            )
+        if self.save_checkpoints:
+            callbacks.append(
+                ModelCheckpoint(
+                    save_last=True,
+                    save_top_k=1,
+                    dirpath=self.infra.folder,
+                    filename=self.checkpoint_filename,
+                    monitor=self.monitor,
+                    mode=mode,
+                    save_on_train_epoch_end=True,
+                )
+            )
+
+        trainer = pl.Trainer(
+            strategy="auto" if self.infra.gpus_per_node == 1 else "fsdp",
+            devices=override_n_devices or self.infra.gpus_per_node,
+            accelerator=self.accelerator,
+            max_epochs=self.n_epochs,
+            max_steps=self.max_steps,
+            limit_train_batches=self.limit_train_batches,
+            enable_progress_bar=self.enable_progress_bar,
+            log_every_n_steps=self.log_every_n_steps,
+            fast_dev_run=self.fast_dev_run,
+            callbacks=callbacks,
+            logger=self._logger,
+            enable_checkpointing=self.save_checkpoints,
+            accumulate_grad_batches=self.accumulate_grad_batches,
+        )
+        self._trainer = trainer
+        return trainer
+
+    def fit(self, train_loader: DataLoader, valid_loader: DataLoader) -> None:
+        self._trainer.fit(
+            model=self._model,
+            train_dataloaders=train_loader,
+            val_dataloaders=valid_loader,
+            ckpt_path=self._get_checkpoint_path(),
+        )
+
+    def test(self, test_loader: DataLoader) -> None:
+        if self.checkpoint_path:
+            ckpt_path = self.checkpoint_path
+        else:
+            if self.save_checkpoints:
+                ckpt_path = Path(self.infra.folder) / "best.ckpt"
+            else:
+                ckpt_path = None
+        self._trainer.test(
+            self._model,
+            dataloaders=test_loader,
+            ckpt_path=ckpt_path,
+        )
+
+    def setup_run(self):
+
+        if self.infra.cluster and self.infra.status() != "not submitted":
+            for out_type in ["stdout", "stderr"]:
+                old_path = Path(getattr(self.infra.job().paths, out_type))
+                new_path = Path(self.infra.folder) / f"log.{out_type}"
+                try:
+                    if new_path.exists():
+                        os.remove(new_path)
+                    os.symlink(
+                        old_path,
+                        new_path,
+                    )
+                except Exception:
+                    pass
+        config_path = Path(self.infra.folder) / "config.yaml"
+        os.makedirs(self.infra.folder, exist_ok=True)
+        with open(config_path, "w") as outfile:
+            yaml.dump(
+                self.model_dump(),
+                outfile,
+                indent=4,
+                default_flow_style=False,
+                sort_keys=False,
+            )
+
+    @infra.apply
+    def run(self):
+        import lightning.pytorch as pl
+
+        self.setup_run()
+        self._logger = (
+            self.wandb_config.build(
+                save_dir=self.infra.folder,
+                xp_config=self.model_dump(),
+                id=f"{self.wandb_config.group}-{self.infra.uid().split('-')[-1]}",
+            )
+            if self.wandb_config
+            else None
+        )
+
+        if self.seed is not None:
+            pl.seed_everything(self.seed, workers=True)
+            np.random.seed(self.seed)
+            torch.manual_seed(self.seed)
+
+        loaders = self.data.get_loaders(
+            split_to_build="val" if self.test_only else None
+        )
+        self._setup_trainer(next(iter(loaders.values())))
+
+        if not self.test_only:
+            self.fit(loaders["train"], loaders["val"])
+
+        self.test(loaders["val"])

tribev2/model.py

@@ -0,0 +1,234 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+import typing as tp
+
+import torch
+from einops import rearrange
+from neuralset.dataloader import SegmentData
+from neuraltrain.models.base import BaseModelConfig
+from neuraltrain.models.common import Mlp, SubjectLayers, SubjectLayersModel
+from neuraltrain.models.transformer import TransformerEncoder
+from torch import nn
+
+logger = logging.getLogger(__name__)
+
+
+class TemporalSmoothing(BaseModelConfig):
+    kernel_size: int = 9
+    sigma: float | None = None
+
+    def build(self, dim: int) -> nn.Module:
+
+        def gaussian_kernel_1d(kernel_size: int, sigma: float):
+            x = torch.arange(kernel_size) - kernel_size // 2
+            kernel = torch.exp(-0.5 * (x / sigma) ** 2)
+            kernel = kernel / kernel.sum()
+            return kernel.view(1, 1, -1)
+
+        conv = nn.Conv1d(
+            dim,
+            dim,
+            kernel_size=self.kernel_size,
+            padding=self.kernel_size // 2,
+            bias=False,
+            groups=dim,
+        )
+        if self.sigma is not None:
+            kernel = gaussian_kernel_1d(kernel_size=self.kernel_size, sigma=self.sigma)
+            kernel = kernel.repeat(dim, 1, 1)
+            conv.weight.data = kernel
+            conv.requires_grad = False
+        return conv
+
+
+class FmriEncoder(BaseModelConfig):
+
+    # architecture
+    projector: BaseModelConfig = Mlp(norm_layer="layer", activation_layer="gelu")
+    combiner: Mlp | None = Mlp(norm_layer="layer", activation_layer="gelu")
+    encoder: TransformerEncoder | None = TransformerEncoder()
+    # other hyperparameters
+    time_pos_embedding: bool = True
+    subject_embedding: bool = False
+    subject_layers: SubjectLayers | None = SubjectLayers()
+    hidden: int = 256
+    max_seq_len: int = 1024
+    dropout: float = 0.0
+    extractor_aggregation: tp.Literal["stack", "sum", "cat"] = "cat"
+    layer_aggregation: tp.Literal["mean", "cat"] = "cat"
+    linear_baseline: bool = False
+    modality_dropout: float = 0.0
+    temporal_dropout: float = 0.0
+    low_rank_head: int | None = None
+    temporal_smoothing: TemporalSmoothing | None = None
+
+    def model_post_init(self, __context):
+        if self.encoder is not None:
+            for key in ["attn_dropout", "ff_dropout", "layer_dropout"]:
+                setattr(self.encoder, key, self.dropout)
+        if hasattr(self.projector, "dropout"):
+            self.projector.dropout = self.dropout
+        return super().model_post_init(__context)
+
+    def build(
+        self, feature_dims: dict[int], n_outputs: int, n_output_timesteps: int
+    ) -> nn.Module:
+        return FmriEncoderModel(
+            feature_dims,
+            n_outputs,
+            n_output_timesteps,
+            config=self,
+        )
+
+
+class FmriEncoderModel(nn.Module):
+
+    def __init__(
+        self,
+        feature_dims: dict[str, tuple[int, int]],
+        n_outputs: int,
+        n_output_timesteps: int,
+        config: FmriEncoder,
+    ):
+        super().__init__()
+        self.config = config
+        self.feature_dims = feature_dims
+        self.n_outputs = n_outputs
+        self.n_output_timesteps = n_output_timesteps
+        self.projectors = nn.ModuleDict()
+        self.pooler = nn.AdaptiveAvgPool1d(n_output_timesteps)
+        hidden = config.hidden
+        for modality, tup in feature_dims.items():
+            if tup is None:
+                logger.warning(
+                    "%s has no feature dimensions. Skipping projector.", modality
+                )
+                continue
+            else:
+                num_layers, feature_dim = tup
+            input_dim = (
+                feature_dim * num_layers
+                if config.layer_aggregation == "cat"
+                else feature_dim
+            )
+            output_dim = (
+                hidden // len(feature_dims)
+                if config.extractor_aggregation == "cat"
+                else hidden
+            )
+            self.projectors[modality] = self.config.projector.build(
+                input_dim, output_dim
+            )
+        input_dim = (
+            (hidden // len(feature_dims)) * len(feature_dims)
+            if config.extractor_aggregation == "cat"
+            else hidden
+        )
+        if self.config.combiner is not None:
+            self.combiner = self.config.combiner.build(input_dim, hidden)
+        else:
+            assert (
+                hidden % len(feature_dims) == 0
+            ), "hidden must be divisible by the number of modalities if there is no combiner"
+            self.combiner = nn.Identity()
+        if config.low_rank_head is not None:
+            self.low_rank_head = nn.Linear(hidden, config.low_rank_head, bias=False)
+            bottleneck = config.low_rank_head
+        else:
+            bottleneck = hidden
+        self.predictor = config.subject_layers.build(
+            in_channels=bottleneck,
+            out_channels=n_outputs,
+        )
+        if config.temporal_smoothing is not None:
+            self.temporal_smoothing = config.temporal_smoothing.build(dim=hidden)
+        if not config.linear_baseline:
+            if config.time_pos_embedding:
+                self.time_pos_embed = nn.Parameter(
+                    torch.randn(1, config.max_seq_len, hidden)
+                )
+            if config.subject_embedding:
+                self.subject_embed = nn.Embedding(config.n_subjects, hidden)
+            self.encoder = config.encoder.build(dim=hidden)
+
+    @property
+    def device(self) -> torch.device:
+        return next(self.parameters()).device
+
+    def forward(self, batch: SegmentData, pool_outputs: bool = True) -> torch.Tensor:
+        x = self.aggregate_features(batch)  # B, T, H
+        subject_id = batch.data.get("subject_id", None)
+        if hasattr(self, "temporal_smoothing"):
+            x = self.temporal_smoothing(x.transpose(1, 2)).transpose(1, 2)
+        if not self.config.linear_baseline:
+            x = self.transformer_forward(x, subject_id)
+        x = x.transpose(1, 2)  # B, H, T
+        if self.config.low_rank_head is not None:
+            x = self.low_rank_head(x.transpose(1, 2)).transpose(1, 2)
+        x = self.predictor(x, subject_id)  # B, O, T
+        if pool_outputs:
+            out = self.pooler(x)  # B, O, T'
+        else:
+            out = x
+        return out
+
+    def aggregate_features(self, batch):
+        tensors = []
+        # get B, T
+        for modality in batch.data.keys():
+            if modality in self.feature_dims:
+                break
+        x = batch.data[modality]
+        B, T = x.shape[0], x.shape[-1]
+        for modality in self.feature_dims.keys():
+            if modality not in self.projectors or modality not in batch.data:
+                data = torch.zeros(
+                    B, T, self.config.hidden // len(self.feature_dims)
+                ).to(x.device)
+            else:
+                data = batch.data[modality]  # B, L, H, T
+                data = data.to(torch.float32)
+                if data.ndim == 3:
+                    data = data.unsqueeze(1)
+                # mean over layers
+                if self.config.layer_aggregation == "mean":
+                    data = data.mean(dim=1)
+                elif self.config.layer_aggregation == "cat":
+                    data = rearrange(data, "b l d t -> b (l d) t")
+                data = data.transpose(1, 2)
+                assert data.ndim == 3  # B, T, D
+                if isinstance(self.projectors[modality], SubjectLayersModel):
+                    data = self.projectors[modality](
+                        data.transpose(1, 2), batch.data["subject_id"]
+                    ).transpose(1, 2)
+                else:
+                    data = self.projectors[modality](data)  # B, T, H
+                if self.config.modality_dropout > 0 and self.training:
+                    mask = torch.rand(data.shape[0]) < self.config.modality_dropout
+                    data[mask, :] = torch.zeros_like(data[mask, :])
+            tensors.append(data)
+        if self.config.extractor_aggregation == "stack":
+            out = torch.cat(tensors, dim=1)
+        elif self.config.extractor_aggregation == "cat":
+            out = torch.cat(tensors, dim=-1)
+        elif self.config.extractor_aggregation == "sum":
+            out = sum(tensors)
+        if self.config.temporal_dropout > 0 and self.training:
+            for batch_idx in range(out.shape[0]):
+                mask = torch.rand(out.shape[1]) < self.config.temporal_dropout
+                out[batch_idx, mask, :] = torch.zeros_like(out[batch_idx, mask, :])
+        return out
+
+    def transformer_forward(self, x, subject_id=None):
+        x = self.combiner(x)
+        if hasattr(self, "time_pos_embed"):
+            x = x + self.time_pos_embed[:, : x.size(1)]
+        if hasattr(self, "subject_embed"):
+            x = x + self.subject_embed(subject_id)
+        x = self.encoder(x)
+        return x

tribev2/pl_module.py

@@ -0,0 +1,155 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Custom lightning module that wraps a pytorch model.
+"""
+
+import typing as tp
+from pathlib import Path
+
+import lightning.pytorch as pl
+from einops import rearrange
+from neuralset.dataloader import SegmentData
+from neuraltrain.optimizers import BaseOptimizer
+from torch import nn
+from torchmetrics import Metric
+
+
+class BrainModule(pl.LightningModule):
+    """Torch-lightning module for fMRI encoding model training."""
+
+    def __init__(
+        self,
+        model: nn.Module,
+        loss: nn.Module,
+        optim_config: BaseOptimizer,
+        metrics: dict[str, Metric],
+        checkpoint_path: Path | None = None,
+        config: dict[str, tp.Any] | None = None,
+    ) -> None:
+        super().__init__()
+        self.model = model
+        self.checkpoint_path = checkpoint_path
+        self.config = config
+
+        # Optimizer
+        self.optim_config = optim_config
+
+        self.loss = loss
+        self.metrics = metrics
+
+    def forward(self, batch):
+        return self.model(batch)
+
+    def on_save_checkpoint(self, checkpoint):
+        checkpoint["model_build_args"] = {
+            "feature_dims": self.model.feature_dims,
+            "n_outputs": self.model.n_outputs,
+            "n_output_timesteps": self.model.n_output_timesteps,
+        }
+
+    def _run_step(
+        self, batch: SegmentData, batch_idx, step_name, dataloader_idx: int = 0
+    ):
+        y_true = batch.data["fmri"]  # B, D, T
+        y_pred = self.forward(batch)  # B, D, T
+        if step_name == "val":
+            y_true = y_true[:, :, self.config["data.overlap_trs_val"] :]
+            y_pred = y_pred[:, :, self.config["data.overlap_trs_val"] :]
+        subject_ids_flat = batch.data["subject_id"].repeat_interleave(
+            y_pred.shape[2], 0
+        )
+
+        y_pred_flat = rearrange(y_pred, "b d t -> (b t) d")
+        y_true_flat = rearrange(y_true, "b d t -> (b t) d")
+        if not self.config["data.stride_drop_incomplete"]:
+            bad_indices = (y_true_flat == 0).all(dim=1)
+            y_pred_flat = y_pred_flat[~bad_indices]
+            y_true_flat = y_true_flat[~bad_indices]
+            subject_ids_flat = subject_ids_flat[~bad_indices]
+
+        loss = self.loss(y_pred_flat, y_true_flat).mean()
+        log_kwargs = {
+            "on_step": True if step_name == "train" else False,
+            "on_epoch": True,
+            "logger": True,
+            "prog_bar": True,
+            "batch_size": y_pred.shape[0],
+        }
+
+        self.log(
+            f"{step_name}/loss",
+            loss,
+            **log_kwargs,
+        )
+
+        # Compute metrics
+        for metric_name, metric in self.metrics.items():
+            if metric_name.startswith(step_name):
+                if "grouped" in metric.__class__.__name__.lower():
+                    metric.update(y_pred_flat, y_true_flat, groups=subject_ids_flat)
+                else:
+                    if "retrieval" in metric_name:
+                        metric.update(y_pred.mean(dim=-1), y_true.mean(dim=-1))
+                    else:
+                        metric.update(y_pred_flat, y_true_flat)
+                    self.log(
+                        metric_name,
+                        metric,
+                        **log_kwargs,
+                    )
+        return loss, y_pred.detach().cpu(), y_true.detach().cpu()
+
+    def on_val_or_test_epoch_end(self, step_name: str) -> None:
+        for metric_name, metric in self.metrics.items():
+            if metric_name.startswith(step_name):
+                if "grouped" in metric.__class__.__name__.lower():
+                    subject_id_to_name = {
+                        v: k
+                        for k, v in self.config[
+                            "data.subject_id.predefined_mapping"
+                        ].items()
+                    }
+                    metric_dict = {
+                        metric_name + "/" + subject_id_to_name[int(k)]: v
+                        for k, v in metric.compute().items()
+                    }
+                    self.log_dict(metric_dict)
+                    metric.reset()
+
+    def on_validation_epoch_end(self) -> None:
+        self.on_val_or_test_epoch_end("val")
+        return super().on_validation_epoch_end()
+
+    def on_test_epoch_end(self) -> None:
+        self.on_val_or_test_epoch_end("test")
+        return super().on_test_epoch_end()
+
+    def training_step(self, batch: SegmentData, batch_idx):
+        loss, _, _ = self._run_step(batch, batch_idx, step_name="train")
+        return loss
+
+    def validation_step(self, batch: SegmentData, batch_idx, dataloader_idx: int = 0):
+        _, y_pred, y_true = self._run_step(
+            batch, batch_idx, step_name="val", dataloader_idx=dataloader_idx
+        )
+        return y_pred, y_true
+
+    def test_step(self, batch: SegmentData, batch_idx, dataloader_idx: int = 0):
+        _, y_pred, y_true = self._run_step(
+            batch, batch_idx, step_name="test", dataloader_idx=dataloader_idx
+        )
+        return y_pred, y_true
+
+    def configure_optimizers(self):
+        optim_config = self.optim_config.copy()
+        unfrozen_params = [p for p in self.parameters() if p.requires_grad]
+        if self.config["max_steps"] > 0:
+            total_steps = self.config["max_steps"]
+        else:
+            total_steps = self.trainer.estimated_stepping_batches
+        optimizer = optim_config.build(unfrozen_params, total_steps=total_steps)
+        return optimizer

tribev2/plotting/__init__.py

@@ -0,0 +1,26 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .base import BasePlotBrain
+from .cortical import PlotBrainNilearn
+from .cortical_pv import PlotBrainPyvista
+from .subcortical import get_subcortical_roi_indices, plot_subcortical
+from .utils import (
+    combine_mosaics,
+    convert_ax_to_2d,
+    convert_ax_to_3d,
+    get_cmap,
+    get_pval_stars,
+    label_ax,
+    move_ax,
+    plot_colorbar,
+    plot_rgb_colorbar,
+    saturate_colors,
+    set_title,
+    shrink_ax,
+)
+
+PlotBrain = PlotBrainPyvista

tribev2/plotting/base.py

@@ -0,0 +1,497 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import typing as tp
+from functools import lru_cache
+
+import matplotlib
+import nibabel as nib
+import numpy as np
+import pydantic
+from neuralset.extractors.neuro import FSAVERAGE_SIZES
+from nilearn import datasets, image, maskers, surface
+from scipy.spatial import cKDTree
+
+cached_fetch_surf_fsaverage = lru_cache(datasets.fetch_surf_fsaverage)
+
+
+class BasePlotBrain(pydantic.BaseModel):
+    mesh: (
+        tp.Literal["fsaverage3", "fsaverage4", "fsaverage5", "fsaverage6", "fsaverage7"]
+        | None
+    ) = "fsaverage5"
+    inflate: bool | tp.Literal["half"] = "half"
+    bg_map: tp.Literal["sulcal", "curvature", "thresholded"] = "sulcal"
+    hemisphere_gap: float = 0
+    atlas_name: str | None = None
+    atlas_dim: int | None = None
+    vol_to_surf_kwargs: dict | None = None
+    model_config = pydantic.ConfigDict(extra="forbid")
+
+    VIEW_DICT: tp.ClassVar[dict] = {}
+
+    def model_post_init(self, __context: tp.Any) -> None:
+        self._mesh = self.get_mesh()
+
+    # ------------------------------------------------------------------
+    # Axes helpers
+    # ------------------------------------------------------------------
+
+    def get_axarr_and_views(self, axes, views):
+        if isinstance(axes, dict):
+            axes = {k: self._convert_ax(ax) for k, ax in axes.items()}
+            if all(k in self.VIEW_DICT for k in axes):
+                views, axarr = zip(*axes.items())
+            else:
+                axarr = list(axes.values())
+        elif isinstance(axes, (list, np.ndarray)):
+            axarr = axes
+        elif isinstance(axes, matplotlib.axes.Axes):
+            axarr = [axes]
+        assert len(views) == len(
+            axarr
+        ), f"Number of views and axes must match, got {len(views)} and {len(axarr)}"
+        return views, axarr
+
+    def _convert_ax(self, ax):
+        """Hook for subclasses that need to convert axes (e.g. 3D -> 2D)."""
+        return ax
+
+    # ------------------------------------------------------------------
+    # Atlas / volume-to-surface helpers
+    # ------------------------------------------------------------------
+
+    def get_atlas(self):
+        if not hasattr(self, "_atlas"):
+            if self.atlas_name == "schaefer_2018":
+                atlas = datasets.fetch_atlas_schaefer_2018(n_rois=self.atlas_dim)
+            elif self.atlas_name == "difumo":
+                atlas = datasets.fetch_atlas_difumo(dimension=self.atlas_dim)
+            self._atlas = atlas
+        return self._atlas
+
+    @property
+    def atlas_masker(self):
+        if not hasattr(self, "_atlas_masker"):
+            atlas = self.get_atlas()
+            if self.atlas_name == "schaefer_2018":
+                atlas_masker = maskers.NiftiLabelsMasker(labels_img=atlas["maps"])
+            elif self.atlas_name == "difumo":
+                atlas_masker = maskers.NiftiMapsMasker(maps_img=atlas["maps"])
+            atlas_masker.fit()
+            self._atlas_masker = atlas_masker
+        return self._atlas_masker
+
+    def atlas_to_surf(self, signals, img_threshold: float | None = None):
+        signals_nii = self.signals_to_nii(signals)
+        return self.vol_to_surf(signals_nii, img_threshold=img_threshold)
+
+    def vol_to_surf(self, signals_nii, img_threshold: float | None = None):
+        vol_to_surf_kwargs = self.vol_to_surf_kwargs or {}
+        if img_threshold is not None:
+            signals_nii = image.threshold_img(
+                signals_nii,
+                threshold=img_threshold,
+                copy=False,
+                copy_header=True,
+            )
+        fsaverage = cached_fetch_surf_fsaverage(mesh=self.mesh)
+        hemis = [
+            surface.vol_to_surf(
+                signals_nii,
+                surf_mesh=fsaverage[f"pial_{hemi}"],
+                kind="ball",
+                **vol_to_surf_kwargs,
+            )
+            for hemi in ("left", "right")
+        ]
+        return np.concatenate(hemis)
+
+    def signals_to_nii(self, signals):
+        out = self.atlas_masker.inverse_transform(signals)
+        if isinstance(self.atlas_masker, maskers.NiftiMapsMasker):
+            data = out.get_fdata()
+            lo, hi = signals.min(), signals.max()
+            data = (data - data.min()) / (data.max() - data.min())
+            data = data * (hi - lo) + lo
+            out = nib.Nifti1Image(data, out.affine, out.header)
+        return out
+
+    # ------------------------------------------------------------------
+    # Mesh loading (eager – called once in model_post_init)
+    # ------------------------------------------------------------------
+
+    def get_mesh(self) -> dict:
+        """Load mesh geometry and background maps for both hemispheres.
+
+        Returns a dict with keys ``'left'``, ``'right'``, ``'both'``,
+        each mapping to ``{'coords': array, 'faces': array, 'bg_map': array}``.
+        The ``'both'`` entry has hemisphere_gap applied.
+        """
+        fs_out = cached_fetch_surf_fsaverage(self.mesh)
+
+        out = {}
+        for hemi in ("left", "right"):
+            infl_out_xyz, _ = nib.load(getattr(fs_out, f"infl_{hemi}")).darrays
+            pial_xyz, faces = nib.load(getattr(fs_out, f"pial_{hemi}")).darrays
+
+            alpha = 0.5
+            jr_xyz = infl_out_xyz.data * alpha + (1 - alpha) * pial_xyz.data
+            if self.inflate == "half":
+                coords = jr_xyz
+            elif self.inflate is True:
+                coords = infl_out_xyz.data
+            elif self.inflate is False:
+                coords = pial_xyz.data
+
+            bg_key = "curv" if self.bg_map == "curvature" else "sulc"
+            bg_map = nib.load(getattr(fs_out, f"{bg_key}_{hemi}")).darrays[0].data
+            if self.bg_map == "thresholded":
+                bg_map = 1.0 * (bg_map > -0.10)
+                bg_map[-1] = -5
+                bg_map[-2] = 2.0
+            if hemi == "left":
+                coords[:, 0] = coords[:, 0] - coords[:, 0].max() - self.hemisphere_gap
+            else:
+                coords[:, 0] = coords[:, 0] - coords[:, 0].min() + self.hemisphere_gap
+
+            out[hemi] = dict(coords=coords, faces=faces.data, bg_map=bg_map)
+
+        out["both"] = dict(
+            coords=np.r_[out["left"]["coords"], out["right"]["coords"]],
+            faces=np.r_[
+                out["left"]["faces"],
+                out["right"]["faces"] + out["left"]["faces"].max() + 1,
+            ],
+            bg_map=np.r_[out["left"]["bg_map"], out["right"]["bg_map"]],
+        )
+
+        return out
+
+    # ------------------------------------------------------------------
+    # Stat-map upsampling (lazy – called per data array)
+    # ------------------------------------------------------------------
+
+    def get_stat_map(self, data: np.ndarray) -> dict:
+        """Split vertex data into hemispheres, upsampling if needed.
+
+        Returns ``{'left': array, 'right': array, 'both': array}``.
+        """
+        in_mesh = None
+        for name, size in FSAVERAGE_SIZES.items():
+            if data.shape[0] // 2 == size:
+                in_mesh = name
+                break
+        if in_mesh is None:
+            raise ValueError(f"Incoherent number of vertices: {data.shape[0]}")
+
+        left = data[: len(data) // 2]
+        right = data[len(data) // 2 :]
+
+        if in_mesh != self.mesh:
+            fs_in = cached_fetch_surf_fsaverage(in_mesh)
+            fs_out = cached_fetch_surf_fsaverage(self.mesh)
+            resampled = {}
+            for hemi, values in (("left", left), ("right", right)):
+                infl_in_xyz, _ = nib.load(getattr(fs_in, f"infl_{hemi}")).darrays
+                infl_out_xyz, _ = nib.load(getattr(fs_out, f"infl_{hemi}")).darrays
+                tree = cKDTree(infl_in_xyz.data)
+                distances, indices = tree.query(infl_out_xyz.data, k=5)
+                if "int" in data.dtype.name:
+                    # get most frequent
+                    resampled[hemi] = np.apply_along_axis(
+                        lambda x: np.bincount(x).argmax(), axis=1, arr=values[indices]
+                    )
+                else:
+                    distances = np.where(distances == 0, 1e-12, distances)
+                    weights = 1 / distances
+                    weights = weights / weights.sum(axis=1, keepdims=True)
+                    resampled[hemi] = np.sum(values[indices] * weights, axis=1)
+            left, right = resampled["left"], resampled["right"]
+
+        return dict(left=left, right=right, both=np.r_[left, right])
+
+    def get_hemis(self, data: np.ndarray) -> dict:
+        """Convenience: combine ``self._mesh`` geometry with stat-map data."""
+        stat_maps = self.get_stat_map(data)
+        out = {}
+        for hemi in ("left", "right", "both"):
+            m = self._mesh[hemi]
+            out[hemi] = dict(
+                stat_map=stat_maps[hemi],
+                surf_mesh=(m["coords"], m["faces"]),
+                bg_map=m["bg_map"],
+                hemi=hemi,
+            )
+        return out
+
+    # ------------------------------------------------------------------
+    # Multi-timestep plotting
+    # ------------------------------------------------------------------
+
+    def plot_timesteps(
+        self,
+        neuro: np.ndarray | dict[str, np.ndarray],
+        segments=None,
+        *,
+        plot_every_k_timesteps: int = 1,
+        trues=None,
+        norm_percentile=None,
+        show_stimuli=False,
+        views: str | dict[str, str] = "left",
+        timestamps: list[float] | None = None,
+        **kwargs,
+    ):
+        import matplotlib.pyplot as plt
+        from tqdm import tqdm
+
+        from tribev2.plotting.utils import robust_normalize
+
+        TEXT_KEY, SOUND_KEY, VIDEO_KEY = "Text", "Audio", "Video"
+
+        if isinstance(neuro, np.ndarray):
+            neuro = {"Brain reponse": neuro}
+        assert all(
+            v.ndim == 2 for v in neuro.values()
+        ), "Neuro must be a dictionary of 2D arrays"
+        if isinstance(views, dict):
+            assert all(
+                key in views.keys() for key in neuro.keys()
+            ), f"Views keys {views.keys()} do not match neuro keys {neuro.keys()}"
+        total_n_timesteps = len(list(neuro.values())[0])
+        assert (
+            total_n_timesteps % plot_every_k_timesteps == 0
+        ), f"Total number of timesteps {total_n_timesteps} must be divisible by plot_every_k_timesteps {plot_every_k_timesteps}"
+        neuro = {k: v[::plot_every_k_timesteps] for k, v in neuro.items()}
+        n_timesteps = len(list(neuro.values())[0])
+        if timestamps is None:
+            timestamps = range(
+                0, n_timesteps * plot_every_k_timesteps, plot_every_k_timesteps
+            )
+        else:
+            assert (
+                len(timestamps) == n_timesteps
+            ), f"Number of timestamps {len(timestamps)} must match number of timesteps {n_timesteps}"
+        if norm_percentile is not None:
+            neuro = {
+                k: robust_normalize(v, percentile=norm_percentile)
+                for k, v in neuro.items()
+            }
+
+        mosaic = [[f"{k}_{i}" for i in range(n_timesteps)] for k in neuro]
+        height_ratios = [1 for _ in neuro]
+        if show_stimuli:
+            from tribev2.plotting.utils import get_clip
+
+            has_image = any(get_clip(segment) is not None for segment in segments)
+            stimuli_mosaic = [
+                [SOUND_KEY] * n_timesteps,
+                [TEXT_KEY] * n_timesteps,
+            ]
+            stimuli_height_ratios = [0.3, 0.3]
+            if has_image:
+                stimuli_mosaic = [
+                    [f"{VIDEO_KEY}_{i}" for i in range(n_timesteps)]
+                ] + stimuli_mosaic
+                stimuli_height_ratios = [0.7] + stimuli_height_ratios
+            mosaic = stimuli_mosaic + mosaic
+            height_ratios = stimuli_height_ratios + height_ratios
+
+        fig, axes = plt.subplot_mosaic(
+            mosaic,
+            height_ratios=height_ratios,
+            figsize=(2.5 * n_timesteps, 2 * sum(height_ratios)),
+            gridspec_kw={"wspace": 0.0, "hspace": 0},
+        )
+        for k, ax in axes.items():
+            if (
+                k.startswith(TEXT_KEY)
+                or k.startswith(SOUND_KEY)
+                or k.startswith(VIDEO_KEY)
+            ):
+                fig.delaxes(ax)
+                axes[k] = fig.add_subplot(ax.get_subplotspec())
+
+        for i in tqdm(range(n_timesteps), desc="Plotting..."):
+            for j, (key, value) in enumerate(neuro.items()):
+                self.plot_surf(
+                    value[i],
+                    axes=axes[f"{key}_{i}"],
+                    views=views[key] if isinstance(views, dict) else views,
+                    **kwargs,
+                )
+                if j == len(neuro) - 1:
+                    title = (
+                        f"t={timestamps[i]}s" if timestamps is not None else f"t={i}s"
+                    )
+                    fig.text(
+                        0.5,
+                        -0.1,
+                        title,
+                        transform=axes[f"{key}_{i}"].transAxes,
+                        ha="center",
+                        va="center",
+                    )
+
+        if show_stimuli:
+            self.plot_stimuli(
+                segments, axes, plot_every_k_timesteps=plot_every_k_timesteps
+            )
+
+        first_neuro_keys = [key + "_0" for key in list(neuro.keys())]
+        left, full_width = (
+            axes[first_neuro_keys[0]].get_position().x0,
+            fig.get_figwidth(),
+        )
+        for key, label in zip(
+            first_neuro_keys + [TEXT_KEY, SOUND_KEY, f"{VIDEO_KEY}_0"],
+            list(neuro.keys()) + [TEXT_KEY, SOUND_KEY, VIDEO_KEY],
+        ):
+            if key not in axes:
+                continue
+            pos = axes[key].get_position()
+            fig.text(
+                left,
+                (pos.y0 + pos.y1) / 2,
+                label + "\n\n\n",
+                rotation="vertical",
+                va="center",
+                ha="center",
+                transform=fig.transFigure,
+            )
+        return fig
+
+    @staticmethod
+    def plot_stimuli(
+        segments,
+        axes,
+        plot_every_k_timesteps=1,
+    ):
+        import matplotlib.pyplot as plt
+
+        from tribev2.plotting.utils import get_audio, get_clip
+
+        TEXT_KEY, SOUND_KEY, VIDEO_KEY = "Text", "Audio", "Video"
+
+        audio = get_audio(
+            segments[0], stop_offset=(len(segments) - 1) * segments[0].duration
+        )
+        soundarray = audio.to_soundarray().mean(axis=1)
+        axes[SOUND_KEY].plot(soundarray, color="k")
+        axes[SOUND_KEY].set_xlim(0, len(soundarray))
+        axes[SOUND_KEY].axis("off")
+        axes[TEXT_KEY].axis("off")
+        full_start, full_duration = (
+            segments[0].start,
+            len(segments) * segments[0].duration,
+        )
+
+        for i, segment in enumerate(segments):
+            if f"{VIDEO_KEY}_0" in axes and i % plot_every_k_timesteps == 0:
+                ax_idx = i // plot_every_k_timesteps
+                img = get_clip(segment).get_frame(0)
+                margin = img.shape[1] * 0.0
+                ax = axes[f"{VIDEO_KEY}_{ax_idx}"]
+                im = ax.imshow(img)
+                patch = plt.matplotlib.patches.FancyBboxPatch(
+                    (0, 0),
+                    img.shape[1],
+                    img.shape[0],
+                    boxstyle="round,pad=0,rounding_size=200",
+                    transform=ax.transData,
+                    clip_on=False,
+                    facecolor="none",
+                    edgecolor="none",
+                )
+                ax.add_patch(patch)
+                im.set_clip_path(patch)
+                ax.set_xlim(-margin, img.shape[1] + margin)
+                ax.set_ylim(img.shape[0] + margin, -margin)
+                ax.axis("off")
+            events = segment.events
+            words = events[events.type == "Word"]
+            for word in words.itertuples():
+                if word.start < full_start:
+                    continue
+                axes[TEXT_KEY].text(
+                    (word.start - full_start) / full_duration,
+                    0.5,
+                    word.text,
+                    color="k",
+                    transform=axes[TEXT_KEY].transAxes,
+                    ha="center",
+                    va="center",
+                    rotation=45,
+                    fontsize=10,
+                )
+
+    def plot_timesteps_mp4(
+        self,
+        neuro,
+        filepath,
+        *,
+        segments=None,
+        suptitle=None,
+        interpolated_fps=None,
+        norm_percentile=100,
+        **plot_kwargs,
+    ):
+        import subprocess
+        from pathlib import Path
+
+        import matplotlib.pyplot as plt
+        from tqdm import tqdm
+
+        filepath = Path(filepath)
+        tmp_dir = filepath.parent / "tmp"
+        tmp_dir.mkdir(parents=True, exist_ok=True)
+        for i in tqdm(range(len(neuro)), desc="Plotting..."):
+            out_fig, ax = plt.subplots(1, 1, figsize=(3, 3))
+            self.plot_surf(
+                neuro[i],
+                axes=[ax],
+                **plot_kwargs,
+            )
+            title = suptitle or f"t = {i}s"
+            out_fig.suptitle(title, fontsize=14, fontweight="bold")
+            if segments:
+                from tribev2.plotting.utils import get_text
+
+                words = " ".join(get_text(segments[i]).split(" ")[-8:])
+                out_fig.text(0.1, 0.92, words, fontsize=9, ha="left", va="top")
+            tmp_fig = tmp_dir / f"tmp_{i:05d}.png"
+            out_fig.savefig(tmp_fig, dpi=300)
+            plt.close(out_fig)
+        cmd = [
+            "ffmpeg",
+            "-y",
+            "-framerate",
+            str(1),
+            "-i",
+            f"{str(tmp_dir)}/tmp_%05d.png",
+        ]
+        if interpolated_fps is not None:
+            cmd.append("-vf")
+            cmd.append(f"minterpolate=fps={interpolated_fps}")
+        cmd.extend(
+            [
+                "-c:v",
+                "libx264",
+                "-crf",
+                "18",
+                "-pix_fmt",
+                "yuv420p",
+                str(filepath),
+            ]
+        )
+        subprocess.run(cmd)
+
+    # ------------------------------------------------------------------
+    # Rendering (subclasses must implement)
+    # ------------------------------------------------------------------
+
+    def plot_surf(self, *args, **kwargs):
+        raise NotImplementedError

tribev2/plotting/cortical.py

@@ -0,0 +1,311 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import typing as tp
+from pathlib import Path
+
+import matplotlib
+import matplotlib.pyplot as plt
+import numpy as np
+from neuralset.extractors.neuro import FSAVERAGE_SIZES
+from nilearn.datasets import load_fsaverage
+from nilearn.plotting import plot_surf_roi, plot_surf_stat_map
+
+from tribev2.utils import get_hcp_roi_indices
+
+from .base import BasePlotBrain
+from .utils import get_cmap, get_scalar_mappable, robust_normalize, saturate_colors
+
+VIEW_DICT = {
+    "left": (0, 180),
+    "right": (0, 0),
+    "medial_left": (0, 0),
+    "medial_right": (0, 180),
+    "dorsal": (90, 0),
+    "ventral": (-90, 0),
+    "anterior": (0, 90),
+    "posterior": (0, -90),
+    "posterior_left": (0, -135),
+    "posterior_right": (0, -45),
+    "posterior_ventral": (-45, -90),
+    "posterior_ventral_left": (-10, -135),
+}
+
+
+class PlotBrainNilearn(BasePlotBrain):
+
+    VIEW_DICT: tp.ClassVar[dict] = VIEW_DICT
+
+    def get_fig_axes(self, views):
+        if isinstance(views, str):
+            views = [views]
+        n_rows, n_cols = (1, len(views)) if len(views) <= 4 else (2, len(views) // 2)
+        fig, axarr = plt.subplots(
+            n_rows,
+            n_cols,
+            figsize=(2 * n_cols, 2 * n_rows),
+            subplot_kw={"projection": "3d"},
+            gridspec_kw={"wspace": 0, "hspace": -0.2},
+        )
+        if len(views) == 1:
+            axarr = [axarr]
+        else:
+            axarr = axarr.flatten()
+        return fig, axarr
+
+    def plot_surf(
+        self,
+        signals: np.ndarray,
+        norm_percentile=None,
+        colorbar_title: str | None = None,
+        alpha_cmap: tp.Tuple[float, float] | None = None,
+        axes: tp.Any | None = None,
+        colorbar_kwargs: dict | None = None,
+        views: str | list[str] | list[tuple[int, int]] = "left",
+        annotated_rois: list[str] | None = None,
+        vmin: float | None = None,
+        vmax: float | None = None,
+        symmetric_cbar: bool = False,
+        threshold: float | None = None,
+        cmap: str = "hot",
+        colorbar: bool = False,
+    ):
+        if isinstance(views, str):
+            views = [views]
+        if axes is None:
+            fig, axarr = self.get_fig_axes(views=views)
+        else:
+            views, axarr = self.get_axarr_and_views(axes, views)
+            fig = None
+
+        if self.atlas_name is not None:
+            signals = self.atlas_to_surf(signals)
+        elif signals.ndim == 3:
+            signals = self.vol_to_surf(signals)
+        assert (
+            signals.shape[0] // 2 in FSAVERAGE_SIZES.values()
+        ), f"Incoherent number of vertices: {signals.shape[0]}"
+        if norm_percentile is not None:
+            signals = robust_normalize(signals, percentile=norm_percentile)
+        hemis = self.get_hemis(signals)
+        if str(signals.dtype).startswith("int"):
+            plot_fn = plot_surf_roi
+            for k in hemis:
+                hemis[k]["roi_map"] = hemis[k].pop("stat_map")
+            sm = None
+        else:
+            plot_fn = plot_surf_stat_map
+            cmap = get_cmap(cmap, alpha_cmap=alpha_cmap)
+            sm = get_scalar_mappable(
+                signals,
+                cmap,
+                vmin=vmin,
+                vmax=vmax,
+                threshold=threshold,
+                symmetric_cbar=symmetric_cbar,
+            )
+        for i, (view, ax) in enumerate(zip(views, axarr)):
+            selected_hemi = (
+                "left"
+                if view in ["left", "medial_left"]
+                else "right" if view in ["right", "medial_right"] else "both"
+            )
+            if isinstance(view, str):
+                view = VIEW_DICT[view]
+            plot_kwargs = {
+                "axes": ax,
+                "view": view,
+                "figure": fig,
+                "bg_on_data": (
+                    False
+                    if (alpha_cmap is not None or plot_fn == plot_surf_roi)
+                    else True
+                ),
+                "cmap": cmap,
+                "vmin": vmin,
+                "vmax": vmax,
+                "threshold": threshold,
+                "colorbar": False,
+            }
+            if plot_fn == plot_surf_stat_map:
+                plot_kwargs["symmetric_cbar"] = symmetric_cbar
+            plot_fn(**hemis[selected_hemi], **plot_kwargs)
+            if annotated_rois is not None:
+                self.annotate_rois(ax, annotated_rois, hemi=selected_hemi)
+            ax.set_box_aspect(None, zoom=1.4)
+
+        if colorbar:
+            if fig is None:
+                cbar = plt.colorbar(
+                    sm,
+                    format="{x:0.2f}",
+                    label=colorbar_title,
+                    ax=axarr[-1],
+                    **colorbar_kwargs if colorbar_kwargs is not None else {},
+                    shrink=0.5,
+                )
+            else:
+                cb_ax = fig.add_axes([0.9, 0.2, 0.02, 0.6])
+                cbar = fig.colorbar(
+                    sm,
+                    format="{x:0.2f}",
+                    label=colorbar_title,
+                    cax=cb_ax,
+                    **colorbar_kwargs if colorbar_kwargs is not None else {},
+                )
+        return sm
+
+    def plot_surf_rgb(
+        self,
+        signals: tp.List[np.ndarray],
+        alpha_signals: np.ndarray | None = None,
+        norm_percentile=95,
+        alpha_bg=0,
+        cmap: tp.Literal["rgb", "rgb_argmax", "tab10"] = "rgb",
+        saturation_factor: None | float = None,
+        save_path: str | None = None,
+        axes: tp.List[matplotlib.axes.Axes] | None = None,
+        views: list[str] | list[tuple[int, int]] = ["left"],
+        bg_on_data=False,
+    ):
+        if isinstance(views, str):
+            views = [views]
+        if axes is None:
+            fig, axarr = self.get_fig_axes(views=views)
+        else:
+            views, axarr = self.get_axarr_and_views(axes, views)
+            fig = None
+
+        fsaverage_meshes = load_fsaverage(mesh=self.mesh)
+        if self.atlas_name is not None:
+            signals = [self.atlas_to_surf(signal) for signal in signals]
+        elif signals[0].ndim == 4:
+            signals = [self.vol_to_surf(signal) for signal in signals]
+        for signal in signals:
+            assert (
+                signal.shape[0] // 2 in FSAVERAGE_SIZES.values()
+            ), f"Incoherent number of vertices: {signal.shape[0]//2}"
+        hemis = [self.get_hemis(signal) for signal in signals]
+        if alpha_signals is not None:
+            alpha_hemis = self.get_hemis(alpha_signals)
+        data = dict()
+        for selected_hemis in ("left", "right", "both"):
+            vertices, faces = hemis[0][selected_hemis]["surf_mesh"]
+            colors = np.stack(
+                [hemi[selected_hemis]["stat_map"] for hemi in hemis], axis=1
+            )
+            if cmap.startswith("rgb"):
+                if len(signals) == 2:
+                    colors = np.concatenate(
+                        [colors, np.zeros((colors.shape[0], 1))], axis=1
+                    )
+                assert colors.shape[1] == 3
+                if "argmax" in cmap:
+                    colors = robust_normalize(colors, axis=1, percentile=100)
+                    func = np.vectorize(lambda color: 0 if color < 1 else 1)
+                    colors = func(colors)
+                if norm_percentile is not None:
+                    colors = robust_normalize(
+                        colors, percentile=norm_percentile, two_sided=False
+                    )
+                if saturation_factor is not None:
+                    colors = saturate_colors(colors, saturation_factor)
+                colors = np.concatenate([colors, np.ones((colors.shape[0], 1))], axis=1)
+            else:
+                indices = np.argmax(colors, axis=1)
+                cm = get_cmap(cmap)
+                colors = cm(indices - 1)
+                colors[indices == 0, :3] = np.zeros_like(colors[indices == 0, :3])
+            if alpha_signals is not None:
+                alpha = alpha_hemis[selected_hemis]["stat_map"]
+                alpha_bg = 1 - alpha[:, None]
+
+            bg = hemis[0][selected_hemis]["bg_map"]
+            cmap_bg = plt.get_cmap("gray_r")
+            bg = robust_normalize(bg, percentile=100)
+            bg = cmap_bg(bg)
+            if bg_on_data:
+                colors[:, :3] = colors[:, :3] * bg[:, :3]
+            else:
+                colors[:, :3] = colors[:, :3] * (1 - alpha_bg) + bg[:, :3] * alpha_bg
+            face_colors = np.mean(colors[faces], axis=1)
+            data[selected_hemis] = dict(
+                vertex_colors=colors,
+                face_colors=face_colors,
+                vertices=vertices,
+                faces=faces,
+            )
+
+        for view, ax in zip(views, axarr):
+            selected_hemis = (
+                "left" if "left" in view else "right" if "right" in view else "both"
+            )
+            colors = data[selected_hemis]["face_colors"]
+            vertices = data[selected_hemis]["vertices"]
+            faces = data[selected_hemis]["faces"]
+
+            p3dcollec = ax.plot_trisurf(
+                vertices[:, 0],
+                vertices[:, 1],
+                vertices[:, 2],
+                triangles=faces,
+                linewidth=0.1,
+                antialiased=False,
+                color="white",
+            )
+            ax.set_box_aspect(None, zoom=1.4)
+            limits = [vertices.min(), vertices.max()]
+            ax.set_xlim(*limits)
+            ax.set_ylim(*limits)
+            p3dcollec.set_facecolors(colors)
+            ax.set_axis_off()
+            ax.view_init(*VIEW_DICT[view])
+        if save_path is not None:
+            save_path = Path(save_path)
+            save_path.parent.mkdir(parents=True, exist_ok=True)
+            np.save(save_path.with_suffix(".npy"), colors)
+
+        return data["both"]["vertex_colors"]
+
+    def save_gif(self, ax, save_path: str | None = None):
+        import matplotlib.animation as animation
+
+        if save_path is None:
+            save_path = "rgb_animation.gif"
+
+        angles = np.linspace(0, 360, 100, endpoint=False)
+
+        def animate(i):
+            ax.view_init(elev=0, azim=angles[i])
+            return (ax,)
+
+        from matplotlib.animation import FuncAnimation
+
+        ani = FuncAnimation(ax.figure, animate, frames=len(angles), interval=30)
+        writer = animation.PillowWriter(fps=30, bitrate=1800)
+        ani.save(save_path, writer=writer)
+
+    def annotate_rois(
+        self,
+        ax,
+        rois: str | list[str] | dict[str, list[str]],
+        hemi: str = "left",
+        **kwargs,
+    ):
+        if isinstance(rois, str):
+            rois = [rois]
+        assert hemi in ["left", "right"]
+        data = np.zeros(2 * FSAVERAGE_SIZES[self.mesh])
+        vertices = self.get_hemis(data)["both"]["surf_mesh"][0]
+        if hemi == "left":
+            vertices = vertices[: FSAVERAGE_SIZES[self.mesh]]
+        else:
+            vertices = vertices[FSAVERAGE_SIZES[self.mesh] :]
+        for roi in rois:
+            vertex_indices = get_hcp_roi_indices(roi, mesh=self.mesh, hemi=hemi)
+            roi_center = vertices[vertex_indices].mean(axis=0)
+            roi_name = rois[roi] if isinstance(rois, dict) else roi
+            ax.text(roi_center[0], roi_center[1], roi_center[2], roi_name, **kwargs)

tribev2/plotting/cortical_pv.py

@@ -0,0 +1,280 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import tempfile
+import typing as tp
+
+import matplotlib.pyplot as plt
+import numpy as np
+import pyvista as pv
+from neuralset.extractors.neuro import FSAVERAGE_SIZES
+
+from tribev2.utils import get_hcp_roi_indices
+
+from .base import BasePlotBrain
+from .utils import (
+    convert_ax_to_2d,
+    get_cmap,
+    get_scalar_mappable,
+    robust_normalize,
+    saturate_colors,
+    tight_crop,
+)
+
+VIEW_DICT = {
+    "ventral": ([0, 0, -1], [1, 0, 0]),
+    "dorsal": ([0, 0, 1], [0, 1, 0]),
+    "left": ([-1, 0, 0], [0, 0, 1]),
+    "right": ([1, 0, 0], [0, 0, 1]),
+    "anterior": ([0, 1, 0], [0, 0, -1]),
+    "posterior": ([0, -1, 0], [0, 0, 1]),
+    "medial_left": ([1, 0, 0], [0, 0, 1]),
+    "medial_right": ([-1, 0, 0], [0, 0, 1]),
+    "posterior_left": ([-1, 0, 0], [0, 0, 1]),
+    "posterior_right": ([-1, 0, 0], [0, 0, 1]),
+}
+
+
+class PlotBrainPyvista(BasePlotBrain):
+
+    dpi: int = 3000
+    bg_darkness: float = 0
+    ambient: float = 0.3
+    w_pad: float = 0.03
+    h_pad: float = 0.03
+
+    VIEW_DICT: tp.ClassVar[dict] = VIEW_DICT
+
+    def _convert_ax(self, ax):
+        return convert_ax_to_2d(ax)
+
+    def annotate_rois(
+        self,
+        pl: pv.Plotter,
+        rois: str | list[str] | dict[str, str],
+        hemi: str = "left",
+        **kwargs,
+    ):
+        if isinstance(rois, str):
+            rois = [rois]
+        hemis = ["left", "right"] if hemi == "both" else [hemi]
+        n = FSAVERAGE_SIZES[self.mesh]
+        for h in hemis:
+            verts = self._mesh[h]["coords"]
+            for roi in rois:
+                idx = get_hcp_roi_indices(roi, mesh=self.mesh, hemi=h)
+                if h == "right":
+                    idx = np.array(idx) - n
+                center = verts[idx].mean(axis=0)
+                name = rois[roi] if isinstance(rois, dict) else roi
+                pl.add_point_labels(
+                    center.reshape(1, 3),
+                    [name],
+                    shape_opacity=0,
+                    **kwargs,
+                )
+
+    def plot_surf(
+        self,
+        data,
+        axes,
+        views="left",
+        alpha_cmap=None,
+        vmin: float | None = None,
+        vmax: float | None = None,
+        symmetric_cbar: bool = False,
+        threshold: float | None = None,
+        cmap: str = "hot",
+        norm_percentile: float | None = None,
+        annotated_rois: str | list[str] | dict | None = None,
+        annotated_rois_kwargs: dict | None = None,
+    ):
+        if norm_percentile is not None:
+            data = robust_normalize(data, percentile=norm_percentile)
+        if isinstance(views, str):
+            views = [views]
+        views, axes = self.get_axarr_and_views(axes, views)
+        cmap = get_cmap(cmap, alpha_cmap=alpha_cmap)
+        sm = get_scalar_mappable(
+            data,
+            cmap,
+            vmin=vmin,
+            vmax=vmax,
+            threshold=threshold,
+            symmetric_cbar=symmetric_cbar,
+        )
+
+        stat_maps = self.get_stat_map(data)
+
+        for ax, view in zip(axes, views):
+            selected_hemi = (
+                "left"
+                if view in ["left", "medial_left"]
+                else "right" if view in ["right", "medial_right"] else "both"
+            )
+            mesh = self._mesh[selected_hemi]
+            vertices, faces = mesh["coords"], mesh["faces"]
+            stat_map = stat_maps[selected_hemi]
+
+            rgba = sm.to_rgba(stat_map)
+            bg_map = mesh["bg_map"]
+            bg_norm = (bg_map - bg_map.min()) / (bg_map.max() - bg_map.min() + 1e-8)
+            bg_rgb = 1 - np.column_stack(
+                [self.bg_darkness + bg_norm * (1 - self.bg_darkness)] * 3
+            )
+            colors = rgba[:, 3:4] * rgba[:, :3] + (1 - rgba[:, 3:4]) * bg_rgb
+
+            pv_faces = np.column_stack([np.full(len(faces), 3), faces])
+
+            ax_size = ax.get_position()
+            pl = pv.Plotter(
+                window_size=[
+                    int(ax_size.width * self.dpi),
+                    int(ax_size.height * self.dpi),
+                ],
+                off_screen=True,
+            )
+
+            surf = pv.PolyData(vertices, pv_faces)
+            surf.point_data["colors"] = colors
+            pl.add_mesh(
+                surf,
+                scalars="colors",
+                rgb=True,
+                smooth_shading=True,
+                ambient=self.ambient,
+            )
+
+            pl.set_background("white")
+            vec, up = VIEW_DICT[view]
+            pl.view_vector(vec, viewup=up)
+            if annotated_rois is not None:
+                self.annotate_rois(
+                    pl,
+                    annotated_rois,
+                    **(annotated_rois_kwargs or {}),
+                )
+            with tempfile.NamedTemporaryFile(suffix=".png") as tmp:
+                img = pl.screenshot(tmp.name, return_img=True)
+            img = tight_crop(img, w_pad=self.w_pad, h_pad=self.h_pad)
+            pl.clear()
+            ax.axis("off")
+            ax.imshow(img, aspect="equal")
+
+        return sm
+
+    def plot_surf_rgb(
+        self,
+        signals: tp.List[np.ndarray],
+        alpha_signals: np.ndarray | None = None,
+        norm_percentile=95,
+        alpha_bg=0,
+        cmap: tp.Literal["rgb", "rgb_argmax", "tab10"] = "rgb",
+        saturation_factor: None | float = None,
+        axes=None,
+        views: list[str] = ["left"],
+        bg_on_data=False,
+    ):
+        if isinstance(views, str):
+            views = [views]
+        views, axes = self.get_axarr_and_views(axes, views)
+
+        if self.atlas_name is not None:
+            signals = [self.atlas_to_surf(signal) for signal in signals]
+        elif signals[0].ndim == 4:
+            signals = [self.vol_to_surf(signal) for signal in signals]
+
+        hemis = [self.get_hemis(signal) for signal in signals]
+        if alpha_signals is not None:
+            alpha_hemis = self.get_hemis(alpha_signals)
+
+        data = dict()
+        for selected_hemis in ("left", "right", "both"):
+            stat_maps = [hemi[selected_hemis]["stat_map"] for hemi in hemis]
+            colors = np.stack(stat_maps, axis=1)
+
+            if cmap.startswith("rgb"):
+                if len(signals) == 2:
+                    colors = np.concatenate(
+                        [colors, np.zeros((colors.shape[0], 1))], axis=1
+                    )
+                assert colors.shape[1] == 3
+                if "argmax" in cmap:
+                    colors = robust_normalize(colors, axis=1, percentile=100)
+                    colors = (colors >= 1).astype(float)
+                if norm_percentile is not None:
+                    colors = robust_normalize(
+                        colors, percentile=norm_percentile, two_sided=False
+                    )
+                if saturation_factor is not None:
+                    colors = saturate_colors(colors, saturation_factor)
+                colors = np.concatenate([colors, np.ones((colors.shape[0], 1))], axis=1)
+            else:
+                indices = np.argmax(colors, axis=1)
+                cm = get_cmap(cmap)
+                colors = cm(indices - 1)
+                colors[indices == 0, :3] = 0
+
+            if alpha_signals is not None:
+                alpha = alpha_hemis[selected_hemis]["stat_map"]
+                alpha_bg = 1 - alpha[:, None]
+
+            bg = hemis[0][selected_hemis]["bg_map"]
+            cmap_bg = plt.get_cmap("gray_r")
+            bg = robust_normalize(bg, percentile=100)
+            bg = cmap_bg(bg)
+            if bg_on_data:
+                colors[:, :3] = colors[:, :3] * bg[:, :3]
+            else:
+                colors[:, :3] = colors[:, :3] * (1 - alpha_bg) + bg[:, :3] * alpha_bg
+
+            mesh = self._mesh[selected_hemis]
+            data[selected_hemis] = dict(
+                vertex_colors=colors,
+                vertices=mesh["coords"],
+                faces=mesh["faces"],
+            )
+
+        for ax, view in zip(axes, views):
+            selected_hemis = (
+                "left" if "left" in view else "right" if "right" in view else "both"
+            )
+            d = data[selected_hemis]
+
+            pv_faces = np.column_stack([np.full(len(d["faces"]), 3), d["faces"]])
+
+            ax_size = ax.get_position()
+            pl = pv.Plotter(
+                window_size=[
+                    int(ax_size.width * self.dpi),
+                    int(ax_size.height * self.dpi),
+                ],
+                off_screen=True,
+            )
+
+            surf = pv.PolyData(d["vertices"], pv_faces)
+            surf.point_data["colors"] = d["vertex_colors"][:, :3]
+            pl.add_mesh(
+                surf,
+                color="black",
+                scalars="colors",
+                rgb=True,
+                smooth_shading=True,
+                ambient=0.3,
+            )
+
+            vec, up = VIEW_DICT[view]
+            pl.view_vector(vec, viewup=up)
+            with tempfile.NamedTemporaryFile(suffix=".png") as tmp:
+                img = pl.screenshot(
+                    tmp.name, return_img=True, transparent_background=True
+                )
+            img = tight_crop(img, w_pad=self.w_pad, h_pad=self.h_pad)
+            pl.clear()
+            ax.axis("off")
+            ax.imshow(img, aspect="equal")
+
+        return data["both"]["vertex_colors"]

tribev2/plotting/subcortical.py

@@ -0,0 +1,311 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import copy
+import tempfile
+import typing as tp
+from functools import lru_cache
+
+import matplotlib.pyplot as plt
+import nibabel as nib
+import numpy as np
+import pyvista as pv
+import seaborn as sns
+from nilearn import datasets
+from nilearn.surface import vol_to_surf
+from scipy.ndimage import gaussian_filter
+from skimage import measure
+
+from tribev2.plotting.utils import (
+    get_cmap,
+    get_scalar_mappable,
+    robust_normalize,
+    tight_crop,
+)
+
+
+@lru_cache()
+def get_subcortical_mask():
+    atlas = datasets.fetch_atlas_harvard_oxford("sub-maxprob-thr50-2mm")
+    excluded = ["Cortex", "White", "Stem", "Background"]
+    selected_indices = [
+        i
+        for i, label in enumerate(atlas.labels)
+        if any([exc.lower() in label.lower() for exc in excluded])
+    ]
+    mask_data = atlas.maps.get_fdata()
+    mask_data[np.isin(mask_data, selected_indices)] = 0
+    mask = nib.Nifti1Image(mask_data, atlas.maps.affine, atlas.maps.header)
+    return mask
+
+
+def get_subcortical_labels(with_hemi: bool = False):
+    excluded = ["Cortex", "White", "Stem", "Background"]
+    labels = [
+        label
+        for label in cached_ho_atlas().labels
+        if not any([exc.lower() in label.lower() for exc in excluded])
+    ]
+    if not with_hemi:
+        labels = list(
+            set(
+                [
+                    label.replace("Left ", "")
+                    for label in labels
+                    if label.startswith("Left ")
+                ]
+            )
+        )
+    return labels
+
+
+@lru_cache
+def cached_ho_atlas(resolution: tp.Literal["1mm", "2mm"] = "1mm"):
+    return datasets.fetch_atlas_harvard_oxford(f"sub-maxprob-thr50-{resolution}")
+
+
+def get_subcortical_roi_indices(roi: str):
+    subcortical_mask = copy.deepcopy(get_subcortical_mask())
+    data = subcortical_mask.get_fdata()
+    data = data[data > 0]
+    ho_sub = cached_ho_atlas(resolution="2mm")
+    labels = ho_sub.labels
+    sel_labels = [label for label in labels if roi.lower() in label.lower()]
+    assert sel_labels, f"ROI {roi} not found in atlas"
+    sel_indices = [labels.index(label) for label in sel_labels]
+    voxel_indices = np.where(np.isin(data, sel_indices))[0]
+    return voxel_indices
+
+
+def voxel_to_mesh(voxel_scores, label, resolution):
+    subcortical_mask = copy.deepcopy(get_subcortical_mask())
+    data = subcortical_mask.get_fdata()
+    data[data > 0] = voxel_scores
+    nii = nib.Nifti1Image(data, subcortical_mask.affine, subcortical_mask.header)
+    roi_mask = get_mask(label, resolution)
+    mesh = get_mesh(label, resolution)
+    return nii_to_mesh(nii, mesh, mask_img=roi_mask)
+
+
+def nii_to_mesh(nii, mesh, mask_img=None):
+    vertices = mesh.points
+    faces = mesh.faces.reshape(-1, 4)[:, 1:]
+    vertex_vals = vol_to_surf(
+        nii,
+        surf_mesh=(vertices, faces),
+        mask_img=mask_img,
+        kind="line",
+        depth=np.linspace(-3, 0, 40),
+        interpolation="linear",
+    )
+    return vertex_vals
+
+
+@lru_cache()
+def get_mask(label: str, resolution: tp.Literal["1mm", "2mm"] = "1mm"):
+    # fetch Harvard-Oxford subcortical atlas
+    ho_sub = cached_ho_atlas(resolution=resolution)
+    img = ho_sub.maps
+    if label == "Cerebellum":
+        raise NotImplementedError(
+            "Cerebellum atlas (Diedrichsen 2009) is not yet supported. "
+            "Provide the atlas path manually."
+        )
+        img = nib.load(file)
+        mask = img.get_fdata() > 0  # merge all lobules automatically
+    elif label == "Brain-Stem":
+        # subcortical, return hemisphere-specific mesh (default: right)
+        idx = ho_sub.labels.index(label)
+        mask = img.get_fdata() == idx
+    else:
+        if "Left" in label or "Right" in label:
+            idx = ho_sub.labels.index(label)
+            mask = img.get_fdata() == idx
+        else:
+            # merge left + right
+            left_idx = ho_sub.labels.index("Left " + label)
+            right_idx = ho_sub.labels.index("Right " + label)
+            data = img.get_fdata()
+            mask = (data == left_idx) | (data == right_idx)
+
+    nii_mask = nib.Nifti1Image(mask.astype(float), img.affine, img.header)
+
+    return nii_mask
+
+
+@lru_cache()
+def get_mesh(label: str, resolution: tp.Literal["1mm", "2mm"]):
+    """
+    Returns a PyVista mesh for a given label.
+    For 'Cerebellum', 'Cerebral Cortex', and 'Brain-Stem', left and right hemispheres are joined.
+    For other subcortical labels, returns separate left/right meshes.
+    """
+
+    if label == "Cerebral Cortex":
+        fsaverage = datasets.fetch_surf_fsaverage("fsaverage7")
+        nii = nib.load(fsaverage.pial_left)
+        verts = nii.darrays[0].data
+        faces = nii.darrays[1].data
+        faces_pv = np.hstack([np.full((faces.shape[0], 1), 3), faces]).astype(np.int32)
+        mesh = pv.PolyData(verts, faces_pv)
+        return mesh
+
+    nii_mask = get_mask(label, resolution)
+
+    # smooth the mask slightly
+    volume = gaussian_filter(nii_mask.get_fdata().astype(float), sigma=1)
+
+    # marching cubes
+    verts, faces, normals, values = measure.marching_cubes(volume, level=0.9)
+    # Convert voxel coordinates to world/MNI coordinates
+    affine = nii_mask.affine
+    verts = nib.affines.apply_affine(affine, verts)
+
+    # convert faces to PyVista format
+    faces_pv = np.hstack([np.full((faces.shape[0], 1), 3), faces]).astype(np.int32)
+
+    # create PyVista mesh
+    mesh = pv.PolyData(verts, faces_pv)
+
+    # smooth the mesh
+    mesh = mesh.smooth(n_iter=50, relaxation_factor=0.01)
+
+    return mesh
+
+
+def plot_subcortical(
+    ax,
+    *,
+    colors: dict = None,
+    voxel_scores: np.ndarray = None,
+    average_per_roi: bool = False,
+    norm_percentile: int = None,
+    show_cortex: bool = False,
+    show_brain_stem: bool = False,
+    show_cerebellum: bool = False,
+    explode: float = 0.5,
+    resolution: tp.Literal["1mm", "2mm"] = "1mm",
+    show_scalar_bar: bool = False,
+    zoom: float = 1.3,
+    azimuth: float = 15,
+    elevation: float = -10,
+    intensity: float = 1.5,
+    vmin: float | None = None,
+    vmax: float | None = None,
+    symmetric_cbar: bool = False,
+    threshold: float | None = None,
+    cmap: str = "hot",
+    alpha_cmap: tuple[float, float] = None,
+    **plot_kwargs,
+):
+    assert (colors is not None) ^ (
+        voxel_scores is not None
+    ), "Either colors voxel_scores must be provided"
+    labels = get_subcortical_labels(with_hemi=True)
+    if colors is not None:
+        assert isinstance(colors, dict), "Colors must be a dictionary"
+    if voxel_scores is not None:
+        assert voxel_scores.ndim in [1, 2], "voxel_scores must be a 1D or 2D array"
+        if average_per_roi:
+            for label in labels:
+                indices = get_subcortical_roi_indices(label)
+                voxel_scores[indices] = voxel_scores[indices].mean()
+        if norm_percentile:
+            voxel_scores = robust_normalize(voxel_scores, percentile=norm_percentile)
+    if show_cerebellum:
+        labels.append("Cerebellum")
+    if show_cortex:
+        labels.append("Cerebral Cortex")
+    if show_brain_stem:
+        labels.append("Brain-Stem")
+    plotter = pv.Plotter(lighting="none")
+    rgb = False
+    cmap = get_cmap(cmap, alpha_cmap=alpha_cmap)
+    sm = get_scalar_mappable(
+        voxel_scores,
+        cmap,
+        vmin=vmin,
+        vmax=vmax,
+        threshold=threshold,
+        symmetric_cbar=symmetric_cbar,
+    )
+    for label in labels:
+        mesh = get_mesh(label, resolution)
+        if label in ["Cerebral Cortex", "Brain-Stem"]:
+            color = plt.cm.gray(0.8)
+        else:
+            if colors is not None:
+                color = colors[label]
+                scalars = None
+            else:
+                assert voxel_scores is not None
+                color = plt.cm.gray(0.8)
+                if voxel_scores.ndim == 1:
+                    scalars = voxel_to_mesh(voxel_scores, label, resolution)
+                    scalars = sm.to_rgba(scalars)
+                    rgb = True
+                elif voxel_scores.ndim == 2:
+                    assert voxel_scores.shape[0] == 3
+                    scalars = np.stack(
+                        [
+                            voxel_to_mesh(voxel_scores, label, resolution)
+                            for voxel_scores in voxel_scores
+                        ],
+                        axis=1,
+                    )
+                    rgb = True
+        exploded_points = copy.deepcopy(mesh.points)
+        if label == "Cerebral Cortex":
+            exploded_points[:, 0] = (
+                exploded_points[:, 0] + explode * exploded_points[:, 0].mean()
+            )
+        else:
+            exploded_points[:, 2] = (
+                exploded_points[:, 2] + explode * exploded_points.mean(axis=0)[2]
+            )
+        exploded_mesh = pv.PolyData(exploded_points, mesh.faces)
+        plotter.add_mesh(
+            exploded_mesh,
+            color=color,
+            scalars=scalars,
+            rgb=rgb,
+            show_scalar_bar=show_scalar_bar,
+        )
+    plotter.window_size = [300, 300]
+    plotter.camera.zoom(zoom)
+    plotter.camera.azimuth = azimuth
+    plotter.camera.elevation = elevation
+    light = pv.Light(intensity=intensity)
+    light.set_headlight()
+    plotter.add_light(light)
+
+    with tempfile.NamedTemporaryFile(suffix=".png") as tmp:
+        img = plotter.screenshot(tmp.name, return_img=True)
+    img = tight_crop(img)
+    ax.imshow(img)
+    ax.axis("off")
+    return sm
+
+
+if __name__ == "__main__":
+
+    labels = get_subcortical_labels(with_hemi=False)
+    palette = sns.color_palette("Set1", n_colors=len(labels))
+    colors = {
+        f"{hemi} {label}": palette[i]
+        for i, label in enumerate(labels)
+        for hemi in ["Left", "Right"]
+    }
+    plotter = plot_subcortical(
+        colors=colors,
+        average_per_roi=True,
+        cmap="fire",
+        show_cerebellum=False,
+        explode=1,
+        resolution="1mm",
+        zoom=1.3,
+    )
+    plt.show()

tribev2/plotting/utils.py

@@ -0,0 +1,563 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+import re
+from functools import reduce
+
+import colorcet
+import matplotlib
+import matplotlib.pyplot as plt
+import numpy as np
+import seaborn as sns
+from matplotlib.colors import LinearSegmentedColormap
+
+
+def robust_normalize(
+    array, axis=None, percentile=99, clip=True, final_range=None, two_sided=True
+):
+    """Normalize the input array using statistics robust to outliers."""
+    hi = np.percentile(array, percentile, axis=axis, keepdims=True)
+    if two_sided:
+        lo = np.percentile(array, 100 - percentile, axis=axis, keepdims=True)
+    else:
+        lo = np.min(array, axis=axis, keepdims=True)
+    out = (array - lo) / (hi - lo)
+    if clip:
+        out = np.clip(out, 0, 1)
+    if final_range is not None:
+        if final_range == "original":
+            final_range = (lo, hi)
+        out = out * (final_range[1] - final_range[0]) + final_range[0]
+    return out
+
+
+def get_scalar_mappable(
+    data,
+    cmap,
+    vmin=None,
+    vmax=None,
+    symmetric_cbar=False,
+    threshold=None,
+    alpha_cmap=None,
+):
+    vmin = vmin if vmin is not None else np.nanmin(data)
+    vmax = vmax if vmax is not None else np.nanmax(data)
+    if symmetric_cbar:
+        vmin, vmax = -vmax, vmax
+    sm = get_thresholded_sm(vmin, vmax, threshold=threshold, cmap=cmap)
+    return sm
+
+
+def get_thresholded_sm(vmin, vmax, threshold=None, cmap=None):
+
+    if cmap is None:
+        cmap = matplotlib.cm.get_cmap("hot")
+    norm = matplotlib.colors.Normalize(vmin=vmin, vmax=vmax)
+    cmaplist = [cmap(i) for i in range(cmap.N)]
+
+    # set colors to gray for absolute values < threshold
+    if threshold is not None:
+        istart = int(norm(-threshold, clip=True) * (cmap.N - 1))
+        istop = int(norm(threshold, clip=True) * (cmap.N - 1))
+        for i in range(istart, istop):
+            cmaplist[i] = (0.5, 0.5, 0.5, 1.0)
+    our_cmap = LinearSegmentedColormap.from_list("Custom cmap", cmaplist, cmap.N)
+    sm = plt.cm.ScalarMappable(cmap=our_cmap, norm=norm)
+
+    # fake up the array of the scalar mappable.
+    sm._A = []
+
+    return sm
+
+
+def get_pval_stars(pval: float):
+    if pval < 0.0005:
+        return "***"
+    elif pval < 0.005:
+        return "**"
+    elif pval < 0.05:
+        return "*"
+    else:
+        return ""
+
+
+def saturate_colors(rgb: np.ndarray, factor: float):
+    """
+    rgb: tuple/list/array of (R, G, B) in 0-1 range
+    factor: >1 boosts saturation, 1 leaves unchanged, 0 makes gray
+    """
+    rgb = np.array(rgb, dtype=float)
+
+    # Compute luminance (perceptual gray)
+    # Using Rec.709 coefficients for a fairly natural grayscale
+    grayscale_coeffs = np.array([0.2126, 0.7152, 0.0722])
+    if rgb.ndim == 1:
+        lum = np.dot(grayscale_coeffs, rgb)
+    elif rgb.ndim == 2:
+        lum = np.dot(grayscale_coeffs, rgb.T)
+        lum = lum[:, None].repeat(3, axis=1)
+    else:
+        raise ValueError(f"Invalid number of dimensions: {rgb.ndim}")
+
+    # Pull or push the channels relative to gray
+    new_rgb = lum + factor * (rgb - lum)
+
+    # Clamp to 0–1
+    new_rgb = np.clip(new_rgb, 0, 1)
+    return new_rgb
+
+
+def get_alpha_cmap(cmap, threshold: float = 0, scale: float = 1, symmetric=False):
+    """
+    Takes a cmap and makes it transparent below a threshold.
+    Transparency is linearly scaled between threshold and threshold + scale.
+    """
+    assert 0 <= threshold <= 1
+    from matplotlib.colors import ListedColormap
+
+    n_points = 1024
+    new_cmap = cmap(np.linspace(0, 1, n_points))
+    alpha = np.zeros_like(new_cmap[:, 3])
+    # zeros before min, ramp 0 to 1 between min and max, 1 after max
+    min_idx = int(threshold * (n_points - 1))
+    max_idx = int((threshold + scale) * (n_points - 1))
+    ramp = np.linspace(0, 1, max_idx - min_idx)
+    alpha[min_idx : min(max_idx, n_points)] = ramp[: min(max_idx, n_points) - min_idx]
+    alpha[min(max_idx, n_points) :] = 1
+    # alpha[max_idx:] = 1
+    if symmetric:
+        alpha = np.concatenate([alpha[::-2], alpha[::2]])
+    new_cmap[:, 3] = alpha
+    new_cmap = ListedColormap(new_cmap)
+    return new_cmap
+
+
+def get_cmap(
+    cmap_name: str | matplotlib.colors.Colormap,
+    alpha_cmap: tuple[float, float] | None = None,
+):
+    if isinstance(cmap_name, str):
+        cmap = (
+            getattr(matplotlib.cm, cmap_name, None)
+            or getattr(sns.cm, cmap_name, None)
+            or getattr(colorcet.cm, cmap_name, None)
+        )
+    else:
+        cmap = cmap_name
+    if not cmap:
+        raise ValueError(f"Invalid cmap: {cmap}")
+    if alpha_cmap is not None:
+        threshold, scale = alpha_cmap
+        cmap = get_alpha_cmap(
+            cmap,
+            threshold=threshold,
+            scale=scale,
+            symmetric=(cmap_name in ["seismic", "bwr"]),
+        )
+    return cmap
+
+
+def convert_ax_to_3d(ax):
+    if hasattr(ax, "view_init"):
+        return ax
+    pos = ax.get_position()
+    # subplotspec = ax.get_subplotspec()
+    ax3d = ax.figure.add_axes(pos, projection="3d")
+    # ax3d.set_position(pos)
+    ax.remove()
+    return ax3d
+
+
+def convert_ax_to_2d(ax):
+    pos = ax.get_position()
+    ax2d = ax.figure.add_axes(pos)
+    ax.remove()
+    return ax2d
+
+
+def lcm(a, b):
+    return a * b // math.gcd(a, b) if a and b else max(a, b)
+
+
+def _lcm_list(lst):
+    return reduce(lcm, lst, 1)
+
+
+def _repeat_chars(line, times):
+    return "".join(c * times for c in line)
+
+
+def _transpose(block):
+    if not block:
+        return []
+    max_len = max(len(row) for row in block)
+    block = [row.ljust(max_len) for row in block]
+    return ["".join(block[r][c] for r in range(len(block))) for c in range(max_len)]
+
+
+def _check_unique_letters(*blocks):
+    """
+    Ensure all blocks have unique letters across blocks.
+    Raises an AssertionError if any letter appears in more than one block.
+    """
+    unique = set()
+    for i, block in enumerate(blocks, 1):
+        letters = set(block.replace("\n", ""))
+        assert not (
+            letters & unique
+        ), f"Duplicate letters found in block {i}: {letters & unique}"
+        unique.update(letters)
+
+
+def _format_block(mosaic: str) -> str:
+    return mosaic.replace(" ", "").lstrip("\n").rstrip("\n")
+
+
+def combine_mosaics(*blocks, ratio=None, orient="v"):
+
+    if len(blocks) < 2:
+        raise ValueError("Need at least two blocks to combine")
+
+    _check_unique_letters(*blocks)
+    blocks = [_format_block(block) for block in blocks]
+
+    # Normalize input
+    blocks_lines = [block.split("\n") for block in blocks]
+
+    # Normalize ratio
+    if ratio is None:
+        ratios = [1.0] * len(blocks_lines)
+    else:
+        try:
+            ratios = list(ratio)
+            if len(ratios) != len(blocks_lines):
+                raise ValueError
+        except Exception:
+            ratios = [float(ratio)] * len(blocks_lines)
+
+    # Transpose if horizontal
+    transposed = False
+    if orient == "v":
+        blocks_lines = [_transpose(b) for b in blocks_lines]
+        transposed = True
+
+    # Horizontal expansion (columns)
+    cols_list = [max(len(line) for line in b) if b else 0 for b in blocks_lines]
+    Lw = _lcm_list(cols_list)
+    blocks_expanded = []
+    for b, c, r in zip(blocks_lines, cols_list, ratios):
+        b = [line.ljust(c) for line in b]
+        h = max(1, int(round(Lw / c * r)))
+        blocks_expanded.append([_repeat_chars(line, h) for line in b])
+
+    # Vertical expansion (rows)
+    rows_list = [len(b) for b in blocks_expanded]
+    Lh = _lcm_list(rows_list)
+    blocks_tiled = []
+    for b, r in zip(blocks_expanded, ratios):
+        v = max(1, int(round(Lh / len(b))))
+        blocks_tiled.append([line for line in b for _ in range(v)])
+
+    # Combine all blocks
+    combined = ["".join(lines) for lines in zip(*blocks_tiled)]
+
+    # Transpose back if needed
+    if transposed:
+        combined = _transpose(combined)
+
+    return _format_block("\n".join(combined))
+
+
+def plot_colorbar(
+    ax,
+    sm=None,
+    cmap=colorcet.cm.fire,
+    vmin=0,
+    vmax=1,
+    label="R",
+    label_orientation="vertical",
+    orientation="vertical",
+    **kwargs,
+):
+
+    # Hide the axis background, ticks, and spines
+    ax.set_frame_on(False)
+    ax.set_xticks([])
+    ax.set_yticks([])
+
+    # Create a ScalarMappable for the colorbar
+    if sm is None:
+        norm = matplotlib.colors.Normalize(vmin=vmin, vmax=vmax)
+        sm = matplotlib.cm.ScalarMappable(norm=norm, cmap=cmap)
+    sm.set_array([])  # Required for colorbar
+
+    # Draw the colorbar inside the given axis
+    cbar = plt.colorbar(sm, cax=ax, orientation=orientation, **kwargs)
+
+    # Set the label if provided
+    if label is not None:
+        rotation = 0 if label_orientation == "horizontal" else 90
+        cbar.set_label(label, rotation=rotation, labelpad=5)
+
+    # Add border by setting the color and linewidth of all spines
+    rect = matplotlib.patches.Rectangle(
+        (0, 0),
+        1,
+        1,
+        transform=cbar.ax.transAxes,
+        fill=False,
+        edgecolor="k",
+        linewidth=0.5,
+        clip_on=False,
+    )
+    # cbar.ax.add_patch(rect)
+    return cbar
+
+
+def shrink_ax(ax, shrink=0.1, horizontally=True, vertically=True):
+    pos = ax.get_position()
+    # shrink from all sides
+    horizontal_shrink = pos.width * shrink if horizontally else 0
+    vertical_shrink = pos.height * shrink if vertically else 0
+    new_pos = [
+        pos.x0 + horizontal_shrink / 2,
+        pos.y0 + vertical_shrink / 2,
+        pos.width - horizontal_shrink,
+        pos.height - vertical_shrink,
+    ]
+    ax.set_position(new_pos)
+
+
+def move_ax(ax, x=0, y=0):
+    pos = ax.get_position()
+    up = y * pos.height
+    right = x * pos.width
+    new_pos = [
+        pos.x0 + right,
+        pos.y0 + up,
+        pos.width,
+        pos.height,
+    ]
+    ax.set_position(new_pos)
+
+
+def label_ax(
+    ax,
+    label,
+    x_offset=0,
+    y_offset=0.03,
+    fontsize=14,
+    fontweight="bold",
+    facecolor="none",
+    edgecolor="none",
+):
+    pos = ax.get_position()
+    fig = ax.get_figure()
+    fig.text(
+        pos.x0 + x_offset,
+        pos.y1 + y_offset,
+        label,
+        fontsize=fontsize,
+        fontweight=fontweight,
+        ha="center",
+        va="center",
+    )
+
+
+def set_title(axes, title, x_offset=0, y_offset=0, **kwargs):
+    if not isinstance(axes, list):
+        axes = [axes]
+    centers = [(ax.get_position().x0 + ax.get_position().x1) / 2 for ax in axes]
+    x = np.mean(centers)
+    x = x + x_offset
+    y = axes[0].get_position().y1 + y_offset
+    fig = axes[0].get_figure()
+    if not "ha" in kwargs:
+        kwargs["ha"] = "center"
+    if not "va" in kwargs:
+        kwargs["va"] = "top"
+    fig.text(x, y, title, **kwargs)
+
+
+def tight_crop(img, bg_color=(255, 255, 255), tol=5, w_pad=0, h_pad=0):
+    if img.shape[2] == 4:  # alpha channel exists
+        alpha = img[..., 3]
+        ys, xs = np.where(alpha > 0)
+    else:
+        bg = np.array(bg_color)
+        mask = np.any(np.abs(img[..., :3] - bg) > tol, axis=2)
+        ys, xs = np.where(mask)
+
+    if len(xs) == 0:
+        return img  # nothing found
+    left, right, bottom, top = xs.min(), xs.max(), ys.min(), ys.max()
+    w_pad = int(w_pad * (right - left))
+    h_pad = int(h_pad * (top - bottom))
+    left, bottom = max(0, left - w_pad), max(0, bottom - h_pad)
+    right, top = min(img.shape[1], right + w_pad), min(img.shape[0], top + h_pad)
+
+    return img[bottom : top + 1, left : right + 1]
+
+
+def plot_rgb_colorbar(n_cubes=4, alpha=1, labels=["Text", "Audio", "Video"]):
+    # Use a dark background to make the colors pop
+    # plt.style.use('dark_background')
+    fig = plt.figure(figsize=(6, 4))
+    ax = fig.add_subplot(111, projection="3d", proj_type="persp", focal_length=0.15)
+
+    x = np.linspace(0, 1, n_cubes)
+    y = np.linspace(0, 1, n_cubes)
+    z = np.linspace(0, 1, n_cubes)
+    X, Y, Z = np.meshgrid(x, y, z)
+    X, Y, Z = np.ravel(X), np.ravel(Y), np.ravel(Z)
+    colors = np.array([X, Y, Z]).T
+
+    size = 0.2
+
+    for i in range(len(X)):
+        ax.bar3d(
+            X[i] - size / 2,
+            Y[i] - size / 2,
+            Z[i] - size / 2,
+            size,
+            size,
+            size,
+            color=colors[i],
+            alpha=alpha,
+            edgecolor="none",
+        )
+
+    # --- AXIS ARROWS (QUINVERS) ---
+    # We extend the arrows past the data (to 1.4) to show direction clearly
+    arrow_props = dict(arrow_length_ratio=0.1, linewidth=1, pivot="tail")
+    ax.quiver(0, 0, 0, 1.4, 0, 0, color="k", **arrow_props)
+    ax.quiver(0, 0, 0, 0, 1.4, 0, color="k", **arrow_props)
+    ax.quiver(0, 0, 0, 0, 0, 1.4, color="k", **arrow_props)
+
+    # --- LABELS ---
+    # Positioning labels at the tips of the arrows
+    pos = 1.5
+    ax.text(pos, 0, 0, labels[0], color="red", fontweight="bold", ha="center", va="top")
+    ax.text(
+        0, pos, 0, labels[1], color="green", fontweight="bold", ha="center", va="top"
+    )
+    ax.text(0, 0, pos, labels[2], color="blue", fontweight="bold", ha="center")
+
+    # Remove all background clutter
+    ax.set_axis_off()
+    ax.set_facecolor((0, 0, 0, 0))  # Transparent pane
+
+    # view_init: Azimuth -45 degrees keeps the origin cube (black) at the front
+    # ax.view_init(elev=-40, azim=-135)
+    ax.view_init(elev=45, azim=-135 + 180)
+    ax.set_box_aspect(None, zoom=0.85)
+
+    return fig
+
+
+def get_rainbow_brain(mesh="fsaverage5", hemi="both"):
+    import matplotlib.colors as mcolors
+    from nilearn.datasets import fetch_surf_fsaverage
+    from nilearn.surface import load_surf_mesh
+
+    fsaverage = fetch_surf_fsaverage(mesh=mesh)
+    sphere_l, _ = load_surf_mesh(fsaverage["sphere_left"])
+    sphere_r, _ = load_surf_mesh(fsaverage["sphere_right"])
+    if hemi == "both":
+        coords = np.concatenate([sphere_l, sphere_r], axis=0)
+    else:
+        coords = sphere_l if hemi == "left" else sphere_r
+    x, y, z = coords.T
+
+    # SYMMETRY LOGIC:
+    # On fsaverage, +x is Right, -x is Left.
+    # To make them symmetric, we take the absolute value of X
+    # or flip the X for the right hemisphere so that 'lateral' is always
+    # the same direction relative to the color wheel.
+    x_mapped = x if hemi == "left" else -x
+
+    # Hue based on Longitude (using the corrected X)
+    phi = np.arctan2(y, x_mapped)
+    hues = (phi + np.pi) / (2 * np.pi)
+
+    # Value based on Elevation (Z) to make it more distinct
+    # (Optional: adds a slight brightness gradient from bottom to top)
+    z_norm = (z - z.min()) / (z.max() - z.min() + 1e-8)
+    vals = np.clip(0.8 + (z_norm * 0.3), 0, 1)
+
+    hsv = np.stack([hues, np.ones_like(hues) * 0.9, vals], axis=1)
+    return mcolors.hsv_to_rgb(hsv)
+
+
+# ---------------------------------------------------------------------------
+# Segment helpers (moved from analyses/utils.py)
+# ---------------------------------------------------------------------------
+
+
+def has_video(segment) -> bool:
+    return any(e.__class__.__name__ == "Video" for e in segment.ns_events)
+
+
+def has_audio(segment) -> bool:
+    return any(e.__class__.__name__ == "Audio" for e in segment.ns_events)
+
+
+def get_clip(segment, start_offset=0, stop_offset=0):
+    from moviepy import VideoFileClip
+
+    if not has_video(segment):
+        return None
+    video = [e for e in segment.ns_events if e.__class__.__name__ == "Video"][0]
+    clip = VideoFileClip(video.filepath)
+    true_start = video.start - video.offset
+    clip = clip.subclipped(
+        max(segment.start + start_offset - true_start, 0),
+        min(segment.stop + stop_offset - true_start, clip.duration),
+    )
+    return clip
+
+
+def get_audio(segment, start_offset=0, stop_offset=0):
+    from moviepy import AudioFileClip
+
+    if not has_audio(segment):
+        return None
+    audio = [e for e in segment.ns_events if e.__class__.__name__ == "Audio"][0]
+    clip = AudioFileClip(audio.filepath)
+    true_start = audio.start - audio.offset
+    clip = clip.subclipped(
+        max(segment.start + start_offset - true_start, 0),
+        min(segment.stop + stop_offset - true_start, clip.duration),
+    )
+    return clip
+
+
+def get_words(segment, filter=(0, 1), remove_punctuation=True, remove_stopwords=False):
+    start, duration = segment.start, segment.duration
+    clean = (
+        (lambda x: re.sub(r"[^\w\s]", "", x)) if remove_punctuation else (lambda x: x)
+    )
+    words = [
+        clean(e.text.lower())
+        for e in segment.ns_events
+        if e.__class__.__name__ == "Word"
+        and filter[0] <= (e.start - start) / duration <= filter[1]
+    ]
+    if remove_stopwords:
+        from stopwords import get_stopwords
+
+        words = [w for w in words if w not in get_stopwords("english")]
+    return words
+
+
+def get_text(segment, **kwargs) -> str:
+    return " ".join(get_words(segment, **kwargs))
+
+
+if __name__ == "__main__":
+    fig = plot_rgb_colorbar()
+    plt.show()

tribev2/studies/__init__.py

@@ -0,0 +1,10 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from .algonauts2025 import Algonauts2025, Algonauts2025Bold
+from .lahner2024bold import Lahner2024Bold
+from .lebel2023bold import Lebel2023Bold
+from .wen2017 import Wen2017

tribev2/studies/algonauts2025.py

@@ -0,0 +1,315 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+"""Algonauts Project 2025 Challenge: fMRI responses to multimodal movie stimuli.
+
+This study is part of the Algonauts Project 2025 Challenge, using a subset of the
+Courtois NeuroMod dataset (https://www.cneuromod.ca/). Participants watched naturalistic
+video stimuli including episodes from the TV sitcom "Friends" and extractor films while
+undergoing fMRI scanning.
+
+Experimental Design:
+    - 4 participants (sub-01, sub-02, sub-03, sub-05)
+    - Two stimulus types:
+        * "Friends" sitcom: 7 seasons, ~175 episodes, segmented into ~5min chunks (a,b,c,d)
+        * "movie10": 4 extractor films (Bourne, Wolf, Life, Figures) in ~5min chunks
+    - TR = 1.49 seconds
+    - Training data: Friends seasons 1-6, all movies
+    - Test data: Friends season 7
+    - Some movies shown twice (Life, Figures) for reliability analysis
+
+Data Format:
+    - Preprocessed fMRI in MNI152NLin2009cAsym space
+    - Parcellated using Schaefer-1000 atlas (1000 parcels, 7 networks)
+    - HDF5 format
+    - Video stimuli provided as .mkv files
+    - Word-level transcripts with timestamps (.tsv format)
+    - Includes rich multimodal annotations (speech, text, visual extractors)
+
+Download Requirements:
+    - Datalad must be installed (pip install datalad)
+    - Git must be configured
+    - Dataset cloned from: https://github.com/courtois-neuromod/algonauts_2025.competitors.git
+    - Moderate dataset size (~several GB)
+
+Note:
+    This dataset is designed for the Algonauts 2025 Challenge focused on predicting
+    brain responses to complex, naturalistic multimodal stimuli.
+    See: https://algonautsproject.com/2025/index.html
+"""
+
+import ast
+import logging
+import typing as tp
+from itertools import product
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+from neuralset.events import study
+
+logger = logging.getLogger(__name__)
+
+
+class Algonauts2025(study.Study):
+    _SUBJECTS: tp.ClassVar[list[str]] = ["sub-01", "sub-02", "sub-03", "sub-05"]
+    _TASKS: tp.ClassVar[list[str]] = ["friends", "movie10"]
+    _SPACE: tp.ClassVar[str] = "space-MNI152NLin2009cAsym"
+    _ATLAS: tp.ClassVar[str] = "atlas-Schaefer18_parcel-1000Par7Net"
+    _FREQUENCY: tp.ClassVar[float] = 1 / 1.49
+
+    device: tp.ClassVar[str] = "Fmri"
+    dataset_name: tp.ClassVar[str] = "Algonauts 2025 Challenge"
+    url: tp.ClassVar[str] = "https://algonautsproject.com/"
+    bibtex: tp.ClassVar[
+        str
+    ] = """
+    @article{algonauts2025,
+        url = {https://arxiv.org/abs/2501.00504},
+        author = {Gifford,  Alessandro T. and Bersch,  Domenic and St-Laurent,  Marie and Pinsard,  Basile and Boyle,  Julie and Bellec,  Lune and Oliva,  Aude and Roig,  Gemma and Cichy,  Radoslaw M.},
+        keywords = {Neurons and Cognition (q-bio.NC),  FOS: Biological sciences,  FOS: Biological sciences},
+        title = {The Algonauts Project 2025 Challenge: How the Human Brain Makes Sense of Multimodal Movies},
+        publisher = {arXiv},
+        year = {2025},
+        copyright = {Creative Commons Attribution 4.0 International},
+        doi={https://doi.org/10.48550/arXiv.2501.00504},
+        url={https://arxiv.org/abs/2501.00504}
+    }
+    """
+    description: tp.ClassVar[str] = (
+        'Subset of Courtois NeuroMod dataset (boyle2020) with fMRI recordings of subjects watching videos of a popular sitcom ("Friends") for Algonauts 2025'
+    )
+    requirements: tp.ClassVar[tuple[str, ...]] = (
+        "datalad>=0.19.5",
+        "moviepy",
+    )
+
+    _info: tp.ClassVar[study.StudyInfo] = study.StudyInfo(
+        num_timelines=1588,
+        num_subjects=4,
+        num_events_in_query=1700,
+        event_types_in_query={"Fmri", "Video", "Word", "Text"},
+        data_shape=(1000, 592),
+        frequency=0.671,
+        fmri_spaces=("custom",),
+    )
+
+    def _download(self) -> None:
+        raise NotImplementedError("Download method not implemented yet")
+
+    def iter_timelines(self) -> tp.Iterator[dict[str, tp.Any]]:
+        for subject in self._SUBJECTS:
+            for task in self._TASKS:
+                if task == "friends":
+                    season_episode_chunk = range(1, 8), range(1, 26), "abcd"
+                    for season, episode, chunk in product(*season_episode_chunk):
+                        tl = dict(
+                            subject=subject,
+                            task=task,
+                            movie=f"s{season:02d}",
+                            chunk=f"e{episode:02d}{chunk}",
+                            run=0,
+                        )
+                        stim_path = self._get_transcript_filepath(tl)
+                        if (
+                            (season == 5 and episode == 20 and chunk == "a")
+                            or (season == 4 and episode == 1 and chunk == "a")
+                            or (season == 6 and episode == 3 and chunk == "a")
+                            or (season == 4 and episode == 13 and chunk == "b")
+                            or (season == 4 and episode == 1 and chunk == "b")
+                        ):
+                            continue
+                        if stim_path.exists():
+                            yield tl
+                elif task == "movie10":
+                    movie_chunk_run = (
+                        ["bourne", "wolf", "life", "figures"],
+                        range(1, 18),
+                        [1, 2],
+                    )
+                    for movie, chunk, run in product(*movie_chunk_run):  # type: ignore
+                        if movie in ["bourne", "wolf"] and run == 2:
+                            continue
+                        tl = dict(
+                            subject=subject,
+                            task=task,
+                            movie=movie,
+                            chunk=str(chunk),
+                            run=run,
+                        )
+                        stim_path = self._get_transcript_filepath(tl)
+                        if stim_path.exists():
+                            yield tl
+
+    def _get_transcript_filepath(self, timeline: dict[str, tp.Any]) -> Path:
+        tl = timeline
+        base = (
+            self.path
+            / "download/algonauts_2025.competitors/stimuli/transcripts"
+            / tl["task"]
+        )
+        if tl["task"] == "friends":
+            return base / f"s{tl['movie'][-1]}/friends_{tl['movie']}{tl['chunk']}.tsv"
+        elif tl["task"] == "movie10":
+            return (
+                base / f"{tl['movie']}/movie10_{tl['movie']}{int(tl['chunk']):02d}.tsv"
+            )
+        raise ValueError(f"Unknown task: {tl['task']}")
+
+    def _get_movie_filepath(self, timeline: dict[str, tp.Any]) -> Path:
+        tl = timeline
+        base = (
+            self.path
+            / "download/algonauts_2025.competitors/stimuli/movies"
+            / tl["task"]
+        )
+        if tl["task"] == "friends":
+            return base / f"s{tl['movie'][-1]}/friends_{tl['movie']}{tl['chunk']}.mkv"
+        elif tl["task"] == "movie10":
+            return base / f"{tl['movie']}/{tl['movie']}{int(tl['chunk']):02d}.mkv"
+        raise ValueError(f"Unknown task: {tl['task']}")
+
+    def _get_fmri_filepath(self, timeline: dict[str, tp.Any]) -> Path:
+        tl = timeline
+        subj_dir = (
+            self.path
+            / "download/algonauts_2025.competitors/fmri"
+            / tl["subject"]
+            / "func"
+        )
+        stem = f"{tl['subject']}_task-{tl['task']}_{self._SPACE}_{self._ATLAS}"
+        suffix = "_desc-s123456_bold.h5" if tl["task"] == "friends" else "_bold.h5"
+        return subj_dir / f"{stem}{suffix}"
+
+    def _load_fmri(self, timeline: dict[str, tp.Any]) -> tp.Any:
+        import h5py
+
+        tl = timeline
+        fmri_file = self._get_fmri_filepath(timeline)
+        fmri = h5py.File(fmri_file, "r")
+        if tl["task"] == "friends":
+            key = f"{tl['movie'][1:]}{tl['chunk']}"
+        else:
+            key = f"{tl['movie']}{int(tl['chunk']):02d}"
+            if tl["movie"] in ["life", "figures"]:
+                key += f"_run-{tl['run']}"
+        selected_key = [key_ for key_ in fmri.keys() if key in key_]
+        if len(selected_key) != 1:
+            logger.error(
+                "key=%s, selected=%s, available=%s",
+                key,
+                selected_key,
+                list(fmri.keys()),
+            )
+            raise ValueError(f"Multiple or no keys found, {key}, {list(fmri.keys())}")
+        fmri = fmri[selected_key[0]]
+        data = fmri[:].astype(np.float32)
+        import nibabel
+
+        obj = nibabel.Nifti2Image(data.T, affine=np.eye(4))
+        return obj
+
+    def _get_split(self, timeline: dict[str, tp.Any]) -> str:
+        tl = timeline
+        if tl["task"] == "friends":
+            if int(tl["movie"][-1]) in range(1, 7):
+                return "train"
+            elif int(tl["movie"][-1]) == 7:
+                return "test"
+        return "train"
+
+    def _get_fmri_event(self, timeline: dict[str, tp.Any]) -> dict[str, tp.Any]:
+        """Return fmri event dict"""
+        info = study.SpecialLoader(method=self._load_fmri, timeline=timeline).to_json()
+        return dict(type="Fmri", filepath=info, start=0, frequency=self._FREQUENCY)
+
+    def _load_timeline_events(self, timeline: dict[str, tp.Any]) -> pd.DataFrame:
+        all_events = []
+        if (timeline["task"], timeline["movie"]) != ("friends", "s07"):
+            all_events.append(self._get_fmri_event(timeline))
+
+        movie_filepath = self._get_movie_filepath(timeline)
+        movie_event = dict(type="Video", filepath=str(movie_filepath), start=0)
+        all_events.append(movie_event)
+
+        transcript_path = self._get_transcript_filepath(timeline)
+        transcript_df = pd.read_csv(transcript_path, sep="\t")
+        word_events = []
+        for _, row in transcript_df.iterrows():
+            words = ast.literal_eval(row["words_per_tr"])
+            starts = ast.literal_eval(row["onsets_per_tr"])
+            durations = ast.literal_eval(row["durations_per_tr"])
+            for word, start, duration in zip(words, starts, durations):
+                event = dict(
+                    type="Word",
+                    text=word,
+                    start=start,
+                    duration=duration,
+                    stop=start + duration,
+                    language="english",
+                )
+                word_events.append(event)
+        if word_events:
+            word_df = pd.DataFrame(word_events)
+            text = " ".join(word_df["text"].tolist())
+            text_event = dict(
+                type="Text",
+                text=text,
+                start=word_df["start"].min(),
+                duration=word_df["stop"].max() - word_df["start"].min(),
+                stop=word_df["stop"].max(),
+                language="english",
+            )
+            all_events.append(text_event)
+        all_events.extend(word_events)
+
+        events_df = pd.DataFrame(all_events)
+        events_df["split"] = self._get_split(timeline)
+
+        events_df.loc[events_df.type.isin(["Word", "Sentence", "Text"]), "modality"] = (
+            "heard"
+        )
+
+        return events_df
+
+
+class Algonauts2025Bold(Algonauts2025):
+
+    _info: tp.ClassVar[study.StudyInfo] = study.StudyInfo(
+        num_timelines=1588,
+        num_subjects=4,
+        num_events_in_query=1700,
+        event_types_in_query={"Fmri", "Video", "Word", "Text"},
+        data_shape=(76, 90, 71, 592),
+        frequency=0.671,
+        fmri_spaces=("T1w", "MNI152NLin2009cAsym"),
+    )
+
+    def _download(self) -> None:
+        raise NotImplementedError("Download method not implemented yet")
+
+    def _get_fmri_event(self, timeline: dict[str, tp.Any]) -> dict[str, tp.Any]:
+        """Return fmri event dict using fmriprep finder"""
+        tl = timeline
+        if tl["task"] == "friends":
+            task_str = f"{tl['movie']}{tl['chunk']}"
+        else:
+            task_str = f"{tl['movie']}{int(tl['chunk']):02d}"
+        subj_dir = self.path / "download" / f"{tl['task']}.fmriprep" / tl["subject"]
+        task_pattern = f"*_task-{task_str}_*"
+        for session_dir in sorted(subj_dir.iterdir()):
+            if not session_dir.name.startswith("ses-"):
+                continue
+            func_dir = session_dir / "func"
+            if func_dir.exists() and list(func_dir.glob(task_pattern + ".nii.gz")):
+                fp = func_dir / task_pattern
+                return dict(
+                    type="Fmri",
+                    filepath=fp,
+                    layout="fmriprep",
+                    start=0,
+                    frequency=self._FREQUENCY,
+                )
+        raise FileNotFoundError(f"No fMRI file found for {tl}")

tribev2/studies/lahner2024bold.py

@@ -0,0 +1,293 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+"""BOLD Moments: 3T fMRI responses to short naturalistic videos.
+
+This study provides 3T BOLD fMRI data from 10 participants viewing brief (3-second) 
+naturalistic video clips. The dataset is designed to study neural responses to 
+dynamic visual events and includes rich metadata and annotations. The test set's high 
+repetition count (10 reps) enables reliability analysis and within-subject 
+generalization studies.
+
+Experimental Design:
+    - 3T fMRI recordings (TR = 1.75 seconds)
+    - 10 participants
+    - 4 functional scanning sessions per subject (sessions 2-5)
+    - Two sets of stimuli:
+        * Training set: 1,000 unique 3-second video clips (10 runs)
+        * Test set: 102 unique 3-second video clips (3 runs, 10 repetitions each)
+    - Paradigm: passive viewing of naturalistic video clips
+        - Oddball trials included for attention monitoring (excluded from analysis)
+
+Data Format:
+    - BIDS-compliant dataset structure
+    - fMRIPrep preprocessed data (version B recommended by authors)
+    - Available in multiple spaces:
+        * MNI152NLin2009cAsym (volumetric)
+        * T1w (subject-native volumetric)
+        * fsaverage (cortical surface, 163842 vertices per hemisphere)
+        * fsnative (subject-specific cortical surface)
+    - Pre-computed GLM betas available for fsaverage space
+    - Video stimuli
+    - Event annotations:
+        *  LLM-generated captions for middle frames of each video
+
+Download Requirements:
+    - openneuro-py for fMRI data download
+    - Stimuli downloaded from boldmomentsdataset.csail.mit.edu
+    - Moderate dataset size (~several GB)
+    - moviepy required for video processing
+"""
+
+import json
+import pickle as pkl
+import typing as tp
+from pathlib import Path
+
+import nibabel
+import numpy as np
+import pandas as pd
+from neuralset.events import study
+from neuralset.utils import get_bids_filepath, get_masked_bold_image, read_bids_events
+
+
+class Lahner2024Bold(study.Study):
+    device: tp.ClassVar[str] = "Fmri"
+    dataset_name: tp.ClassVar[str] = "BOLD Moments"
+    bibtex: tp.ClassVar[
+        str
+    ] = """
+    @article{Lahner2024,
+        title = {Modeling short visual events through the BOLD moments video fMRI dataset and metadata},
+        volume = {15},
+        ISSN = {2041-1723},
+        url = {http://dx.doi.org/10.1038/s41467-024-50310-3},
+        DOI = {10.1038/s41467-024-50310-3},
+        number = {1},
+        journal = {Nature Communications},
+        publisher = {Springer Science and Business Media LLC},
+        author = {Lahner,  Benjamin and Dwivedi,  Kshitij and Iamshchinina,  Polina and Graumann,  Monika and Lascelles,  Alex and Roig,  Gemma and Gifford,  Alessandro Thomas and Pan,  Bowen and Jin,  SouYoung and Ratan Murty,  N. Apurva and Kay,  Kendrick and Oliva,  Aude and Cichy,  Radoslaw},
+        year = {2024},
+        month = jul 
+    }
+    """
+    licence: tp.ClassVar[str] = "CC0"
+    description: tp.ClassVar[str] = (
+        "BOLD Moments: 3T fMRI from 10 participants viewing 1,000+ brief "
+        "(3-second) naturalistic videos"
+    )
+
+    requirements: tp.ClassVar[tuple[str, ...]] = ("moviepy==2.0.0.dev2",)
+
+    _info: tp.ClassVar[study.StudyInfo] = study.StudyInfo(
+        num_timelines=520,
+        num_subjects=10,
+        num_events_in_query=76,
+        event_types_in_query={"Fmri", "Video"},
+        data_shape=(62, 77, 61, 238),
+        frequency=0.571,
+        fmri_spaces=("custom",),
+    )
+
+    NUM_SUBJECTS: tp.ClassVar[int] = 10
+    NUM_RUNS_PER_SPLIT: tp.ClassVar[dict[str, int]] = {"train": 10, "test": 3}
+
+    DERIVATIVES_FOLDER: tp.ClassVar[str] = "download/derivatives/versionB/fmriprep"
+    SPACES: tp.ClassVar[tuple[str, ...]] = (
+        "MNI152NLin2009cAsym",
+        "T1w",
+        "fsaverage",
+        "fsnative",
+    )
+
+    N_TRIALS_TRAIN: tp.ClassVar[int] = 1000
+    N_TRIALS_TEST: tp.ClassVar[int] = 102
+    N_VOLUMES_TRAIN: tp.ClassVar[int] = 238
+    N_VOLUMES_TEST: tp.ClassVar[int] = 268
+    TR_FMRI_S: tp.ClassVar[float] = 1.75
+
+    def _download(self) -> None:
+        raise NotImplementedError("Download method not implemented yet")
+
+    def _validate_downloaded_data(self) -> None:
+        postfixs = [
+            "_space-MNI152NLin2009cAsym_desc-preproc_bold.nii.gz",
+            "_space-MNI152NLin2009cAsym_desc-brain_mask.nii.gz",
+            "_hemi-R_space-fsaverage_bold.func.gii",
+            "_hemi-L_space-fsaverage_bold.func.gii",
+        ]
+
+        for tl in self.iter_timelines():
+            subj, ses, split, run = tl["subject"], tl["session"], tl["split"], tl["run"]
+            for postfix in postfixs:
+                fp = self.path / (
+                    f"sub-{subj:02d}/ses-{ses:02d}/func/sub-{subj:02d}"
+                    f"_ses-{ses:02d}_task-{split}_run-{run:01d}{postfix}"
+                )
+                if not fp.exists():
+                    msg = f"{fp} is missing. Please download again"
+                    raise RuntimeError(msg)
+
+        for subj in range(1, self.NUM_SUBJECTS + 1):
+            betas_root = (
+                self.path / "download/derivatives/versionB/fsaverage/GLM/"
+                f"sub-{subj:02}/prepared_betas/"
+            )
+            for split in ("train", "test"):
+                for hemi in ("left", "right"):
+                    fp = (
+                        betas_root / f"sub-{subj:02}_organized_betas_task-{split}"
+                        f"_hemi-{hemi}_normalized.pkl"
+                    )
+                    if not fp.exists():
+                        msg = f"{fp} is missing. Please download again"
+                        raise RuntimeError(msg)
+                    with fp.open("rb") as f:
+                        prepared_betas = pkl.load(f)
+                        betas = prepared_betas[0]
+                        n_trials = (
+                            self.N_TRIALS_TEST
+                            if split == "test"
+                            else self.N_TRIALS_TRAIN
+                        )
+                        n_reps = 10 if split == "test" else 3
+                        betas_shape = (n_trials, n_reps, 163842)
+                        if betas.shape != betas_shape:
+                            msg = f"Expected {betas_shape}, got {betas.shape}"
+                            raise RuntimeError(msg)
+                        stims = prepared_betas[1]
+                        if len(stims) != n_trials:
+                            msg = f"Expected {n_trials} stimuli, got {len(stims)}"
+                            raise RuntimeError(msg)
+
+        root = self.path / "stimuli/stimulus_set/stimuli/"
+        for split in ("train", "test"):
+            num_expected = (
+                self.N_TRIALS_TRAIN if split == "train" else self.N_TRIALS_TEST
+            )
+            num_found = len(list((root / split).iterdir()))
+            if num_found != num_expected:
+                msg = f"Expecting {num_expected} stimuli for split {split}"
+                msg += f" but found {num_found}. Please download again"
+                raise RuntimeError(msg)
+
+    def iter_timelines(self) -> tp.Iterator[dict[str, tp.Any]]:
+        for subj in range(1, self.NUM_SUBJECTS + 1):
+            for ses in (2, 3, 4, 5):
+                for split, n_runs in self.NUM_RUNS_PER_SPLIT.items():
+                    for run in range(1, n_runs + 1):
+                        yield dict(subject=subj, session=ses, split=split, run=run)
+
+    def _load_timeline_events(self, timeline: dict[str, tp.Any]) -> pd.DataFrame:
+        tl = dict(timeline)
+        split = tl.pop("split")
+        info = study.SpecialLoader(method=self._load_raw, timeline=timeline).to_json()
+        n_vols = self.N_VOLUMES_TRAIN if split == "train" else self.N_VOLUMES_TEST
+        fmri = {
+            "filepath": info,
+            "type": "Fmri",
+            "start": 0.0,
+            "frequency": 1.0 / self.TR_FMRI_S,
+            "duration": n_vols * self.TR_FMRI_S,
+        }
+        bids_events_df_fp = get_bids_filepath(
+            root_path=self.path / "download",
+            filetype="events",
+            data_type="Fmri",
+            run_padding="01",
+            task=split,
+            **tl,
+        )
+        bids_events_df = read_bids_events(bids_events_df_fp)
+
+        bids_events_df = bids_events_df[bids_events_df.trial_type != "oddball"]
+        ns_events_df = self._get_ns_img_events_df(bids_events_df, timeline)
+        return pd.concat([pd.DataFrame([fmri]), ns_events_df], axis=0)
+
+    def _load_raw(
+        self, timeline: dict[str, tp.Any], space: str = "MNI152NLin2009cAsym"
+    ) -> nibabel.Nifti2Image | nibabel.Nifti1Image:
+        if space in ["MNI152NLin2009cAsym", "T1w"]:
+            return get_masked_bold_image(*self._get_bold_images(timeline, space))
+        elif space in ["fsnative", "fsaverage"]:
+            return self._get_fs(timeline, space)
+        msg = f"{space} is not supported."
+        raise ValueError(msg)
+
+    def _get_ns_img_events_df(
+        self, bids_events_df: pd.DataFrame, timeline: dict[str, tp.Any]
+    ) -> pd.DataFrame:
+        path_to_stimuli = self.path / "stimuli/stimulus_set/stimuli"
+
+        annot_path = (
+            self.path
+            / "download/derivatives/stimuli_metadata/llm_frame_annotations.json"
+        )
+        with annot_path.open("r", encoding="utf8") as f:
+            middle_frame_captions = json.load(f)
+
+        bids_events = bids_events_df.to_dict("records")
+        ns_events = []
+        for bids_event in bids_events:
+            fp = Path(bids_event["stim_file"])
+            filepath = str(path_to_stimuli / fp)
+            captions = "\n".join(next(iter(middle_frame_captions[fp.stem].values())))
+            ns_event = dict(
+                type="Video",
+                start=bids_event["onset"],
+                filepath=filepath,
+                middle_frame_captions=captions,
+            )
+            ns_events.append(ns_event)
+        return pd.DataFrame(ns_events)
+
+    def _get_bold_images(self, timeline: dict[str, tp.Any], space: str):
+        timeline = dict(timeline)
+        timeline["task"] = timeline.pop("split")
+        kwargs = {
+            "root_path": self.path / self.DERIVATIVES_FOLDER,
+            "data_type": "Fmri",
+            "space": space,
+            "run_padding": "01",
+            **timeline,
+        }
+        bold = nibabel.load(get_bids_filepath(**kwargs, filetype="bold"), mmap=True)
+        mask = nibabel.load(
+            get_bids_filepath(**kwargs, filetype="bold_mask"), mmap=True
+        )
+        return (bold, mask)
+
+    def _get_fs(
+        self, timeline: dict[str, tp.Any], space: str = "fsaverage"
+    ) -> nibabel.Nifti2Image:
+        tl = timeline
+        if space not in ["fsaverage", "fsnative"]:
+            msg = f"{space} is not supported. " "Only surfaces 'fsaverage' "
+            msg += "and 'fsnative' are supported for Lahner2024Bold."
+            raise ValueError(msg)
+
+        data = []
+        n_volumes = (
+            self.N_VOLUMES_TRAIN if tl["split"] == "train" else self.N_VOLUMES_TEST
+        )
+        for hemi in ("L", "R"):
+            fp = (
+                self.path
+                / self.DERIVATIVES_FOLDER
+                / f"sub-{int(tl['subject']):02}/ses-{tl['session']:02}"
+                / f"func/sub-{int(tl['subject']):02}_ses-{tl['session']:02}_task-{tl['split']}"
+                f"_run-{tl['run']}_hemi-{hemi}_space-{space}_bold.func.gii"
+            )
+            hemi_data = nibabel.load(fp, mmap=True).darrays  # type: ignore
+            if len(hemi_data) != n_volumes:
+                msg = f"Expected {n_volumes} volumes, got {len(hemi_data)}"
+                raise RuntimeError(msg)
+            if space == "fsaverage" and hemi_data[0].data.shape != (163842,):
+                msg = f"Expected shape (163842,), got {hemi_data[0].data.shape}"
+                raise RuntimeError(msg)
+            np_data = np.stack([darray.data for darray in hemi_data], -1)
+            data.append(np_data)
+        data = np.concatenate(data, axis=0)
+        return nibabel.Nifti2Image(data, np.eye(4))

tribev2/studies/lebel2023bold.py

@@ -0,0 +1,344 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+"""Natural language fMRI dataset: 3T fMRI responses to spoken narrative stories.
+
+This dataset provides fMRI data from participants listening to natural spoken 
+narratives (stories) during 3T scanning. The stimuli include various narrative 
+audio stories with detailed word-level and phoneme-level annotations. The dataset 
+is designed for studying natural language processing in the brain.
+
+Experimental Design:
+    - 3T fMRI recordings (TR = 2.0 seconds)
+    - 8 subjects (UTS01-UTS08)
+    - Subjects 1-3: 82 stories across 20 sessions (extended dataset)
+    - Subjects 4-8: 26-27 stories across 6 sessions
+    - Paradigm: passive listening to naturalistic spoken narratives
+        * Audio narratives with 10-second blank period before story onset
+        * Test story: "wheretheressmoke" (with 10 runs)
+        * Training stories: diverse narrative content
+    - Localizer tasks included: AudioMotorLocalizer, AuditoryLocalizer, 
+      CategoryLocalizer, MotorLocalizer
+
+Data Format:
+    - BIDS-compliant dataset structure (OpenNeuro ds003020)
+    - Two preprocessing versions available (see Study Classes below)
+    - Audio files: WAV format
+    - Event annotations (from TextGrid files)
+        * Word-level timing and text
+        * Phoneme-level timing and text
+        * Audio file paths
+
+Study Classes:
+    1. **Lebel2023Bold**: Uses deepprep preprocessing pipeline
+        - Available spaces: T1w, MNI152NLin6Asym, fsaverage, fsnative
+        - 432 timelines (all sessions/runs)
+        - Full BIDS structure with multiple space outputs
+        
+    2. **LebelProcessed2023Bold**: Uses authors' custom HDF5 preprocessing
+        - Custom cortical surface registration
+        - 200 timelines (aggregated by subject x task)
+        - Data stored in HDF5 format (.hf5 files)
+        - Custom voxel selection and masking
+
+Download Requirements:
+    - OpenNeuro dataset: ds003020
+    - Dataset includes both raw fMRI data and preprocessed derivatives
+    - Audio stimuli (.wav files) and TextGrid annotations included
+    - Deepprep derivatives for Lebel2023Bold
+    - HDF5 preprocessed data for LebelProcessed2023Bold
+    - Python packages:
+        * nltk (v3.8.1) for TextGrid parsing
+        * nltk_contrib (from GitHub) for TextGrid file format
+        * soundfile (>=0.13.1) for audio handling
+        * h5py (>=3.10.0) for HDF5 files (LebelProcessed2023Bold only)
+        * pycortex (for cortical surface visualization, LebelProcessed2023Bold only)
+
+Issues and Considerations:
+    - Subject UTS02: Different scan location and protocol, no localizer data
+    - Subject UTS04: Missing "life.hf5" story scan
+    - Subject UTS05: Low visual acuity, presented auditory cues
+    - UTS01/ses-7/treasureisland: Corrupted NIfTI file, automatically skipped
+    - Preprocessed data has additional 20s removed from beginning
+    - Original preprocessing: https://github.com/HuthLab/deep-fMRI-dataset
+"""
+
+import logging
+import typing as tp
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+from neuralset.events import study
+
+logger = logging.getLogger(__name__)
+
+_DEFAULT_BAD_WORDS = frozenset(
+    [
+        "sentence_start",
+        "sentence_end",
+        "br",
+        "lg",
+        "ls",
+        "ns",
+        "sp",
+        "{BR}",
+        "{LG}",
+        "{LS}",
+        "{NS}",
+        "{SP}",
+    ]
+)
+
+_ANAT_TASKS = [
+    "AudioMotorLocalizer",
+    "AuditoryLocalizer",
+    "CategoryLocalizer",
+    "MotorLocalizer",
+]
+
+SUBJECTS = [f"UTS{i:02d}" for i in range(1, 9)]
+
+
+def _get_audio_file(path: Path | str, task: str) -> Path:
+    path = Path(path)
+    return path / f"stimuli/{task}.wav"
+
+
+def _get_audio_text_file(path: Path | str, task: str) -> Path:
+    path = Path(path)
+    return path / f"derivative/TextGrids/{task}.TextGrid"
+
+
+def _create_audio_events(path: Path | str, task: str) -> list[dict]:
+    events = []
+    dl_path = Path(path)
+    audio_text_file_name = _get_audio_text_file(dl_path, task)
+    audio_wav_file_name = _get_audio_file(dl_path, task)
+
+    split = "train" if task != "wheretheressmoke" else "test"
+
+    events.append(
+        dict(
+            start=0.0,
+            type="Audio",
+            language="english",
+            filepath=str(audio_wav_file_name),
+            split=split,
+        )
+    )
+
+    from nltk_contrib.textgrid import TextGrid
+
+    data = audio_text_file_name.read_text(encoding="utf-8")
+    fid = TextGrid(data)
+
+    for _, tier in enumerate(fid):
+        for recording in tier.simple_transcript:
+            start, stop, text = recording
+            if text != "" and text not in _DEFAULT_BAD_WORDS:
+                if tier.nameid == "phone":
+                    tier_type = "Phoneme"
+                elif tier.nameid == "word":
+                    tier_type = "Word"
+                else:
+                    msg = "Tier must either be phone or word but tier.nameid is %s"
+                    logger.warning(msg, tier.nameid)
+                events.append(
+                    dict(
+                        start=float(start),
+                        text=text.lower(),
+                        duration=float(stop) - float(start),
+                        type=tier_type,
+                        language="english",
+                        filepath=str(audio_wav_file_name),
+                        split=split,
+                    )
+                )
+
+    return events
+
+
+def _get_preprocessed_responses(
+    path: Path | str, task: str, subject: str
+) -> np.ndarray:
+    output = _get_response(Path(path), [task], subject)
+    return output
+
+
+def _get_hf5_path(path: Path | str, subject: str, task: str) -> Path | None:
+    path = Path(path).resolve()
+    hf5_path = path / "derivative" / "preprocessed_data" / subject / f"{task}.hf5"
+    if hf5_path.exists():
+        return hf5_path
+    return None
+
+
+def _get_tasks(path: Path) -> list[str]:
+    path = Path(path).resolve()
+    dl_path = path / "stimuli"
+    tasks = []
+    for fp in dl_path.glob("*.wav"):
+        tasks.append(fp.stem)
+    return tasks
+
+
+def _get_response(path: Path | str, stories, subject) -> np.ndarray:
+    """Get the subject"s fMRI response for stories."""
+    import h5py
+
+    path = Path(path).resolve()
+    base_path = path / f"download/ds003020/derivative/preprocessed_data/{subject}"
+    resp = []
+    for story in stories:
+        resp_path = base_path / f"{story}.hf5"
+        hf = h5py.File(resp_path, "r")
+        resp.extend(hf["data"][:])
+        hf.close()
+    return np.array(resp)
+
+
+class Lebel2023Bold(study.Study):
+    device: tp.ClassVar[str] = "Fmri"
+    licence: tp.ClassVar[str] = "CC0"
+    description: tp.ClassVar[str] = (
+        "Natural language fMRI: 3T fMRI responses from 8 subjects listening to "
+        "spoken narrative stories. Deepprep preprocessing with multiple output spaces "
+        "(T1w, MNI152NLin6Asym, fsaverage, fsnative). 432 timelines with word and "
+        "phoneme-level annotations. Test story: 'wheretheressmoke'."
+    )
+    bibtex: tp.ClassVar[
+        str
+    ] = """
+    @article{lebel2023natural,
+        title={A natural language fMRI dataset for voxelwise encoding models},
+        author={LeBel, Amanda and Wagner, Lauren and Jain, Shailee and Adhikari-Desai, Aneesh and Gupta, Bhavin and Morgenthal, Allyson and Tang, Jerry and Xu, Lixiang and Huth, Alexander G},
+        journal={Scientific Data},
+        volume={10},
+        number={1},
+        pages={555},
+        year={2023},
+        publisher={Nature Publishing Group UK London},
+        doi={https://doi.org/10.1038/s41597-023-02437-z},
+        url={https://www.nature.com/articles/s41597-023-02437-z}
+    }
+
+    @dataset{lebel2023bold,
+        title={A natural language fMRI dataset for voxelwise encoding models},
+        author={LeBel, Amanda and Wagner, Lauren and Jain, Shailee and Adhikari-Desai, Aneesh and
+                Gupta, Bhavin and Morgenthal, Alyssa and Tang, Jerry and Xu, Lixiang and Huth, Alexander G},
+        year={2023},
+        publisher={OpenNeuro},
+        doi={10.18112/openneuro.ds003020.v2.2.0},
+        url={https://openneuro.org/datasets/ds003020}
+    }
+    """
+    requirements: tp.ClassVar[tuple[str, ...]] = (
+        "nltk==3.8.1",
+        "git+https://github.com/nltk/nltk_contrib.git@683961c53f0c122b90fe2d039fe795e0a2b3e997",
+        "soundfile>=0.13.1",
+    )
+    _info: tp.ClassVar[study.StudyInfo] = study.StudyInfo(
+        num_timelines=432,
+        num_subjects=8,
+        num_events_in_query=9199,
+        event_types_in_query={"Fmri", "Audio", "Word", "Phoneme"},
+        data_shape=(57, 65, 56, 363),
+        frequency=0.5,
+        fmri_spaces=("T1w", "MNI152NLin6Asym", "fsaverage", "fsnative"),
+    )
+    TR_FMRI_S: tp.ClassVar[float] = 2.0
+    DERIVATIVES_FOLDER: tp.ClassVar[str] = "download/ds003020-fmriprep"
+
+    def model_post_init(self, __context: tp.Any) -> None:
+        super().model_post_init(__context)
+        self.infra_timelines.version = "v3.4"
+
+    def _download(self) -> None:
+        raise NotImplementedError("Download method not implemented yet")
+
+    def iter_timelines(self) -> tp.Iterator[dict[str, tp.Any]]:
+        """
+        Iterate over the different recording timelines:
+        e.g. subjects x sessions in order with fmri runs
+        """
+        dl_dir = self.path / "download/ds003020"
+        if not dl_dir.exists():
+            raise RuntimeError(f"Missing folder {dl_dir}")
+
+        for subject in SUBJECTS:
+            sessions = 20 if subject in ["UTS01", "UTS02", "UTS03"] else 6
+
+            for sess in range(1, sessions + 1):
+                sess_dir = dl_dir / f"sub-{subject}" / f"ses-{sess}" / "func"
+                tasks = [task.name for task in sess_dir.glob("*_bold.nii.gz")]
+                tasks = sorted({task.split("_")[2].split("-")[1] for task in tasks})
+                for task in tasks:
+                    if task.startswith(tuple(_ANAT_TASKS)):
+                        continue
+                    if subject == "UTS01" and sess == 7 and task == "treasureisland":
+                        msg = "Skipping subject=UTS01, session=7, task=treasureisland as nii.gz is corrupted."
+                        logger.warning(msg)
+                        continue
+
+                    runs = (
+                        list(range(1, 11)) + [None]
+                        if task == "wheretheressmoke"
+                        else [None]
+                    )
+                    for run in runs:
+                        run_infix = f"_run-{run}" if run is not None else ""
+                        filename = f"sub-{subject}_ses-{sess}_task-{task}{run_infix}_bold.nii.gz"
+                        bids_path = sess_dir / filename
+                        if not bids_path.exists():
+                            continue
+
+                        audio_text_file = _get_audio_text_file(path=dl_dir, task=task)
+                        if not audio_text_file.exists():
+                            raise RuntimeError(
+                                f"Missing audio text file: {audio_text_file}"
+                            )
+                        audio_file = _get_audio_file(path=dl_dir, task=task)
+                        if not audio_file.exists():
+                            raise RuntimeError(f"Missing audio file: {audio_file}")
+
+                        yield dict(
+                            subject=subject, session=str(sess), task=task, run=run
+                        )
+
+    def _load_timeline_events(self, timeline: dict[str, tp.Any]) -> pd.DataFrame:
+        """Reads the events of a given timeline"""
+
+        task = timeline["task"]
+        freq = 1.0 / self.TR_FMRI_S
+        events = _create_audio_events(self.path / "download/ds003020", task)
+        subject, session, task, run = (
+            timeline["subject"],
+            timeline["session"],
+            timeline["task"],
+            timeline["run"],
+        )
+        run_substr = f"_run-{run}" if run is not None else ""
+        fp = (
+            self.path
+            / self.DERIVATIVES_FOLDER
+            / f"sub-{subject}/ses-{session}/func"
+            / f"sub-{subject}_ses-{session}_task-{task}{run_substr}_*"
+        )
+        events.append(
+            dict(
+                type="Fmri",
+                start=0.0,
+                filepath=fp,
+                layout="fmriprep",
+                frequency=freq,
+                split="train" if task != "wheretheressmoke" else "test",
+            )
+        )
+        out = pd.DataFrame(events)
+        out.loc[out.type != "Fmri", "start"] += 10
+        out["task"] = task
+        out.loc[out.type != "Fmri", "modality"] = "heard"
+        return out

tribev2/studies/wen2017.py

@@ -0,0 +1,78 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import typing as tp
+from pathlib import Path
+
+import pandas as pd
+from neuralset.events import study
+
+
+def _get_nii_file(path: Path | str, subject: str, seg: str, fmri_run: int) -> Path:
+    path = Path(path)
+    seg_dir = path / subject / "fmri" / seg
+    nii = seg_dir / "mni" / f"{seg}_{fmri_run}_mni.nii.gz"
+    # Outrageously, some test files have a different
+    # naming convention...
+    if not nii.exists():
+        nii = seg_dir / "mni" / f"{seg}_{fmri_run}.mni.nii.gz"
+    assert nii.exists(), f"Missing file {nii} for {subject!r} and {seg!r}"
+    return nii
+
+
+def _get_video_file(path: Path | str, seg: str) -> Path:
+    path = Path(path)
+    return path / f"stimuli/{seg}.mp4"
+
+
+class Wen2017(study.Study):
+    device: tp.ClassVar[str] = "Fmri"
+    licence: tp.ClassVar[str] = "CC-BY 0"
+    url: tp.ClassVar[str] = "https://academic.oup.com/cercor/article/28/12/4136/4560155"
+    TR_FMRI_S: tp.ClassVar[float] = 2.0  # don't rely on nifti header
+
+    def _download(self) -> None:
+        raise NotImplementedError("Download method not implemented yet")
+
+    def iter_timelines(self) -> tp.Iterator[dict[str, tp.Any]]:
+        base = self.path / "download" / "video_fmri_dataset"
+        for subject_dir in base.iterdir():
+            subject = subject_dir.name
+            if not subject.startswith("subject") or not subject_dir.is_dir():
+                continue
+
+            for seg_dir in (subject_dir / "fmri").iterdir():
+                seg = seg_dir.name
+                is_train = seg.startswith("seg")
+                is_test = seg.startswith("test")
+                if not (is_train or is_test):
+                    continue
+                file = _get_video_file(base, seg)
+                if not file.exists():
+                    raise FileNotFoundError(f"Missing video file: {file}")
+
+                fmri_runs = range(1, 3) if is_train else range(1, 11)
+                for run_ in fmri_runs:
+                    nii = _get_nii_file(base, subject, seg, run_)
+                    if not nii.exists():
+                        raise FileNotFoundError(f"Missing nii file: {nii}")
+
+                    yield dict(subject=subject, seg=seg, run=run_)
+
+    def _load_timeline_events(self, timeline: dict[str, tp.Any]) -> pd.DataFrame:
+        import nibabel
+
+        tl = timeline
+        base = self.path / "download" / "video_fmri_dataset"
+        video_file = _get_video_file(base, tl["seg"])
+        nii_file = _get_nii_file(base, tl["subject"], tl["seg"], tl["run"])
+        nii: tp.Any = nibabel.load(nii_file, mmap=True)
+        freq = 1.0 / self.TR_FMRI_S
+        dur = nii.shape[-1] / freq
+        fmri = dict(
+            type="Fmri", start=0, filepath=nii_file, frequency=freq, duration=dur
+        )
+        return pd.DataFrame([dict(type="Video", start=0, filepath=video_file), fmri])

tribev2/utils.py

@@ -0,0 +1,318 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import typing as tp
+from collections import Counter, OrderedDict, defaultdict
+from functools import lru_cache
+from pathlib import Path
+
+import exca
+import mne
+import neuralset as ns
+import numpy as np
+import pandas as pd
+from neuralset.events.study import Chain, Study
+from neuralset.events.transforms import EventsBuilder, EventsTransform
+from neuralset.extractors.neuro import FSAVERAGE_SIZES
+
+from tribev2.eventstransforms import RemoveDuplicates
+
+FMRI_SPACES = {
+    "Algonauts2025Bold": "MNI152NLIN2009C_ASYM_RES_01",
+    "Wen2017": "MNI152NLIN6_ASYM_RES_01",
+    "Lahner2024Bold": "MNI152NLIN2009C_ASYM_RES_01",
+    "Lebel2023Bold": "MNI152NLIN2009C_ASYM_RES_01",
+    "Vanessen2023": "MNI152NLIN6_ASYM_RES_01",
+    "Aliko2020": "MNICOLIN27",
+    "Li2022": "MNICOLIN27",
+    "Nastase2020": "MNI152NLIN2009C_ASYM_RES_01",
+}
+RECORDING_DURATIONS = {
+    "Algonauts2025Bold/sub-01": 66.4,
+    "Algonauts2025Bold/sub-02": 66.4,
+    "Algonauts2025Bold/sub-03": 66.4,
+    "Algonauts2025Bold/sub-04": 0,
+    "Algonauts2025Bold/sub-05": 66.4,
+    "Algonauts2025Bold/sub-06": 0,
+    "Lahner2024Bold/1": 6.2,
+    "Lahner2024Bold/10": 6.2,
+    "Lahner2024Bold/2": 6.2,
+    "Lahner2024Bold/3": 6.2,
+    "Lahner2024Bold/4": 6.2,
+    "Lahner2024Bold/5": 6.2,
+    "Lahner2024Bold/6": 6.2,
+    "Lahner2024Bold/7": 6.2,
+    "Lahner2024Bold/8": 6.2,
+    "Lahner2024Bold/9": 6.2,
+    "Lebel2023Bold/UTS01": 17.9,
+    "Lebel2023Bold/UTS02": 18.1,
+    "Lebel2023Bold/UTS03": 18.1,
+    "Lebel2023Bold/UTS04": 6.2,
+    "Lebel2023Bold/UTS05": 6.4,
+    "Lebel2023Bold/UTS06": 6.4,
+    "Lebel2023Bold/UTS07": 6.4,
+    "Lebel2023Bold/UTS08": 6.4,
+    "Wen2017/subject1": 11.7,
+    "Wen2017/subject2": 11.7,
+    "Wen2017/subject3": 11.7,
+}
+
+
+class MultiStudyLoader(EventsBuilder):
+    """Config for loading multiple studies.
+    Note that the query and enhancers are shared across all studies.
+    For example, setting timeline_index == 0 will select the first timeline of each study.
+    """
+
+    names: str | list[str]
+    path: str | Path
+    transforms: list[EventsTransform] | OrderedDict[str, EventsTransform] | None = None
+    query: str | None = None
+    studies_to_include: list[str] | None = None
+    infra_timelines: exca.MapInfra = exca.MapInfra(cluster="processpool", max_jobs=None)
+
+    def model_post_init(self, log__: tp.Any) -> None:
+        super().model_post_init(log__)
+        if self.studies_to_include is not None:
+            for name in self.studies_to_include:
+                if name not in self.names:
+                    raise ValueError(f"Study {name} not found in {self.names}")
+        self.get_studies()  # run this so that studies are registered (in case _run is cached)
+
+    @infra_timelines.apply(item_uid=str)
+    def dummy(self, items: tp.Iterable[str]) -> tp.Iterator[None]:
+        for item in items:
+            yield None
+
+    def get_studies(self) -> dict[str, Chain]:
+        studies = {}
+        if isinstance(self.names, str):
+            names = [self.names]
+        else:
+            names = self.names
+        for name in names:
+            studies[name] = Study(
+                name=name,
+                path=self.path,
+                query=self.query,
+                infra_timelines=self.infra_timelines,
+            )
+        return studies
+
+    def study_summary(self, apply_query: bool = True) -> pd.DataFrame:
+        summaries = []
+        for name, study in self.get_studies().items():
+            if (
+                apply_query
+                and self.studies_to_include is not None
+                and name not in self.studies_to_include
+            ):
+                continue
+            summary = study.study_summary(apply_query=apply_query)
+            summary.loc[:, "study"] = name
+            summaries.append(summary)
+        return pd.concat(summaries, ignore_index=True)
+
+    def _run(self) -> pd.DataFrame:
+        dfs = []
+        for name, study in self.get_studies().items():
+            if (
+                self.studies_to_include is not None
+                and name not in self.studies_to_include
+            ):
+                continue
+            chain = Chain(steps={"study": study, **OrderedDict(self.transforms)})
+            df = chain.run()
+            df.loc[:, "study"] = name
+            dfs.append(df)
+        out = pd.concat(dfs, ignore_index=True)
+        return out
+
+
+def split_segments_by_time(
+    segments: list[ns.segments.Segment], val_ratio: float, split: str
+) -> list[ns.segments.Segment]:
+    timeline_segments = defaultdict(list)
+    return_segments = []
+    for segment in segments:
+        if len(segment.ns_events) == 0:
+            continue
+        timeline = segment.ns_events[0].timeline
+        timeline_segments[timeline].append(segment)
+    for timeline, segments in timeline_segments.items():
+        start = min(segment.start for segment in segments)
+        stop = max(segment.stop for segment in segments)
+        split_time = start + (stop - start) * val_ratio
+        for segment in segments:
+            if split == "val" and segment.start < split_time:
+                return_segments.append(segment)
+            elif split == "train" and segment.start >= split_time:
+                return_segments.append(segment)
+    return return_segments
+
+
+def assign_fmri_space(events: pd.DataFrame, space: str | None = None) -> pd.DataFrame:
+    assert events.study.nunique() == 1, "Only one study can be assigned at a time"
+    study_name = events.study.unique()[0]
+    if study_name not in FMRI_SPACES:
+        raise ValueError(f"Study {study_name} not found in FMRI_SPACES")
+    default_space = FMRI_SPACES[study_name]
+    assigned_space = space or default_space
+    events.loc[events.type == "Fmri", "space"] = assigned_space
+    return events
+
+
+def set_study_in_average_subject_mode(
+    study: EventsBuilder, trigger_type: str, trigger_field: str = "filepath"
+) -> EventsBuilder:
+    study.transforms["alignevents"] = ns.events.transforms.AlignEvents(
+        trigger_type=trigger_type, trigger_field=trigger_field, types_to_align="Event"
+    )
+    study.transforms["removeduplicates"] = RemoveDuplicates(
+        subset=["start", "stop", "filepath", "type"]
+    )
+    for key in ["chunksounds", "chunkvideos"]:
+        study.transforms.move_to_end(key)
+    return study
+
+
+def get_subject_weights(
+    subject_id_mapping: dict[str, int],
+    weigh_by: tp.Literal[
+        "n_subjects", "speech", "video", "recording_time"
+    ] = "n_subjects",
+) -> dict[str, float]:
+    subject_weights = []
+    if weigh_by in ["speech", "video"]:
+        for subject in subject_id_mapping:
+            if weigh_by == "speech":
+                weight = int(subject.startswith("Lebel"))
+            elif weigh_by == "video":
+                weight = int(subject.startswith("Algonauts"))
+            subject_weights.append(float(weight))
+    elif weigh_by == "recording_time":
+        for subject in subject_id_mapping:
+            if subject not in RECORDING_DURATIONS:
+                raise ValueError(f"Subject {subject} not found in RECORDING_DURATIONS")
+            subject_weights.append(float(RECORDING_DURATIONS[subject]))
+    elif weigh_by == "n_subjects":
+        num_subjects_per_study = Counter(
+            [k.split("/")[0] for k in subject_id_mapping.keys()]
+        )
+        for subject in subject_id_mapping:
+            weight = 1 / num_subjects_per_study[subject.split("/")[0]]
+            subject_weights.append(float(weight))
+    else:
+        raise ValueError(f"Invalid weight type: {weigh_by}")
+    return subject_weights
+
+
+@lru_cache
+def get_hcp_labels(mesh="fsaverage5", combine=False, hemi="both"):
+    """
+    Get the HCP labels for the fsaverage subject.
+    """
+    if hemi in ["right", "left"]:
+        subjects_dir = Path(mne.datasets.sample.data_path()) / "subjects"
+        mne.datasets.fetch_hcp_mmp_parcellation(
+            subjects_dir=subjects_dir, accept=True, verbose=True, combine=combine
+        )
+        name = "HCPMMP1_combined" if combine else "HCPMMP1"
+        with ns.utils.ignore_all():
+            labels = mne.read_labels_from_annot(
+                "fsaverage", name, hemi="both", subjects_dir=subjects_dir
+            )
+        label_to_vertices = {}
+        for label in labels:
+            name, vertices = label.name, np.array(label.vertices)
+            if not combine:
+                name = name[2:]
+            name = name.replace("_ROI", "")  # .replace(" Cortex", "")
+            if (hemi == "right" and "-lh" in name) or (
+                hemi == "left" and "-rh" in name
+            ):
+                continue
+            name = name.replace("-rh", "").replace("-lh", "")
+            label_to_vertices[name] = np.array(vertices)
+        assert sum(len(v) for v in label_to_vertices.values()) == 163842
+        expected_size = FSAVERAGE_SIZES[mesh]
+        index_offset = expected_size if hemi == "right" else 0
+        label_to_vertices = {
+            k: v[v < expected_size] + index_offset for k, v in label_to_vertices.items()
+        }
+        assert sum(len(v) for v in label_to_vertices.values()) == expected_size
+        return label_to_vertices
+    else:
+        assert hemi == "both", f"Invalid hemisphere: {hemi}"
+        left, right = get_hcp_labels(
+            mesh=mesh, combine=combine, hemi="left"
+        ), get_hcp_labels(mesh=mesh, combine=combine, hemi="right")
+        label_to_vertices = {
+            k: np.concatenate([left[k], right[k]]) for k in left.keys()
+        }
+        return label_to_vertices
+
+
+def get_hcp_vertex_labels(mesh="fsaverage5", combine=False):
+    labels = get_hcp_labels(mesh, combine)
+    out = [""] * FSAVERAGE_SIZES[mesh] * 2
+    for label, vertices in labels.items():
+        for vertex in vertices:
+            out[int(vertex)] = label
+    return out
+
+
+def get_hcp_roi_indices(rois: str | list[str], hemi="both", mesh="fsaverage5"):
+    labels = get_hcp_labels(mesh=mesh, combine=False, hemi=hemi)
+    if isinstance(rois, str):
+        rois = [rois]
+    selected_labels = []
+    for roi in rois:
+        if roi[-1] == "*":
+            sel = [label for label in labels.keys() if label.startswith(roi[:-1])]
+        elif roi[0] == "*":
+            sel = [label for label in labels.keys() if label.endswith(roi[1:])]
+        else:
+            sel = [label for label in labels.keys() if label == roi]
+        if not sel:
+            raise ValueError(f"ROI {roi} not found in HCP labels")
+        selected_labels.extend(sel)
+    vertex_indices = np.concatenate([labels[label] for label in selected_labels])
+    return vertex_indices
+
+
+def summarize_by_roi(data: np.ndarray, hemi="both", mesh="fsaverage5"):
+    assert data.ndim == 1, "Data must be 1D"
+    if hemi in ["left", "right", "both"]:
+        labels = get_hcp_labels(mesh=mesh, combine=False, hemi=hemi)
+        out = np.array(
+            [
+                data[get_hcp_roi_indices(roi, hemi=hemi, mesh=mesh)].mean()
+                for roi in labels.keys()
+            ]
+        )
+    elif hemi == "both_separate":
+        out = np.concatenate(
+            [
+                summarize_by_roi(data, hemi="left", mesh=mesh),
+                summarize_by_roi(data, hemi="right", mesh=mesh),
+            ]
+        )
+    else:
+        raise ValueError(f"Invalid hemisphere: {hemi}")
+    return out
+
+
+def get_topk_rois(data: np.ndarray, hemi="both", mesh="fsaverage5", k=10) -> list[str]:
+    values = summarize_by_roi(data, hemi=hemi, mesh=mesh)
+    if hemi == "both_separate":
+        left_labels = get_hcp_labels(mesh=mesh, combine=False, hemi="left").keys()
+        right_labels = get_hcp_labels(mesh=mesh, combine=False, hemi="right").keys()
+        labels = [f"{l}-lh" for l in left_labels] + [f"{l}-rh" for l in right_labels]
+    else:
+        labels = get_hcp_labels(mesh=mesh, combine=False, hemi=hemi).keys()
+    top_k = np.argsort(values)[::-1][:k]
+    return np.array(labels)[top_k]

tribev2/utils_fmri.py

@@ -0,0 +1,248 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import re
+import typing as tp
+from enum import Enum
+
+import neuralset as ns
+import numpy as np
+import pydantic
+from neuralset.extractors.neuro import FSAVERAGE_SIZES
+
+
+class _FmriTemplateSpaceSpec(tp.NamedTuple):
+    id: str
+    shape: tp.Tuple[int, int, int] | None
+
+
+class FmriTemplateSpace(Enum):
+    # MNI - TEMPLATEFLOW (partial)
+    # We keep only 1mm-resolution variants as res mapping is handled by vol_to_surf
+    MNI152LIN_RES_01 = _FmriTemplateSpaceSpec("tpl-MNI152Lin_res-01", (181, 217, 181))
+    MNI152NLIN2009A_ASYM_RES_1 = _FmriTemplateSpaceSpec(
+        "tpl-MNI152NLin2009aAsym_res-1", (197, 233, 189)
+    )
+    MNI152NLIN2009A_SYM_RES_1 = _FmriTemplateSpaceSpec(
+        "tpl-MNI152NLin2009aSym_res-1", (197, 233, 189)
+    )
+    MNI152NLIN2009C_ASYM_RES_01 = _FmriTemplateSpaceSpec(
+        "tpl-MNI152NLin2009cAsym_res-01", (193, 229, 193)
+    )
+    MNI152NLIN2009C_SYM_RES_1 = _FmriTemplateSpaceSpec(
+        "tpl-MNI152NLin2009cSym_res-1", (193, 229, 193)
+    )
+    MNI152NLIN6_ASYM_RES_01 = _FmriTemplateSpaceSpec(
+        "tpl-MNI152NLin6Asym_res-01", (182, 218, 182)
+    )
+    MNI152NLIN6_SYM_RES_01 = _FmriTemplateSpaceSpec(
+        "tpl-MNI152NLin6Asym_res-01", (193, 229, 193)
+    )
+    MNI305 = _FmriTemplateSpaceSpec("tpl-MNI305", (172, 220, 156))
+    MNICOLIN27 = _FmriTemplateSpaceSpec("tpl-MNIColin27", (181, 217, 181))
+
+    # FSAVERAGE
+    FSAVERAGE = _FmriTemplateSpaceSpec("fsaverage", (163842,))
+    FSAVERAGE_6 = _FmriTemplateSpaceSpec("fsaverage6", (40962,))
+    FSAVERAGE_5 = _FmriTemplateSpaceSpec("fsaverage5", (10242,))
+    FSAVERAGE_4 = _FmriTemplateSpaceSpec("fsaverage4", (2562,))
+    FSAVERAGE_3 = _FmriTemplateSpaceSpec("fsaverage3", (642,))
+
+    # CIFTI
+    CIFTI_HCP_FS_LR_32K = _FmriTemplateSpaceSpec("cifti-hcp-fs_LR_32k", (59412,))
+    CIFTI_HCP_FS_LR_164K = _FmriTemplateSpaceSpec("cifti-hcp-fs_LR_164k", (170494,))
+
+    # NATIVE
+    T1W = _FmriTemplateSpaceSpec("T1w", None)
+
+    # OTHER
+    MNI_UNKNOWN = _FmriTemplateSpaceSpec("MNI_unknown", None)  # unknown MNI space
+    UNKNOWN = _FmriTemplateSpaceSpec("unknown", None)  # unknown space
+    CUSTOM = _FmriTemplateSpaceSpec(
+        "custom", None
+    )  # custom space e.g. provided by study authors
+
+
+def is_mni_space(space: FmriTemplateSpace) -> bool:
+    """
+    Check if the given template space is an MNI space.
+    """
+    return space.name.startswith("MNI")
+
+
+def load_mni_mesh(
+    template: FmriTemplateSpace,
+    target_space="fsaverage",
+    base_path: str | None = None,
+) -> dict:
+    """
+    Load MNI surface meshes for both hemispheres and white / pial surfaces.
+
+    Parameters
+    ----------
+    template : FmriTemplateSpace
+    target_space : str
+    base_path : str or None
+        Root directory containing FreeSurfer subjects. If ``None``, reads
+        from the ``FREESURFER_SUBJECTS_DIR`` environment variable.
+
+    Returns
+    -------
+    meshes : dict
+        Dictionary with keys like 'pial_left', 'pial_right', 'white_left', 'white_right'
+        and values as loaded nilearn surface meshes.
+    """
+    import os
+
+    if not re.match(r"^fsaverage[3-6]?$", target_space):
+        raise ValueError(
+            f"target_space must be 'fsaverage' or 'fsaverage3/4/5/6', got '{target_space}'"
+        )
+
+    if not is_mni_space(template):
+        raise ValueError(
+            f"Template {template.value.id} is required to be an MNI space."
+        )
+
+    if base_path is None:
+        base_path = os.getenv("FREESURFER_SUBJECTS_DIR")
+    if base_path is None:
+        raise EnvironmentError(
+            "Set the FREESURFER_SUBJECTS_DIR environment variable to the "
+            "directory containing FreeSurfer subjects, or pass base_path explicitly."
+        )
+
+    from nilearn.surface import load_surf_mesh
+
+    mesh_dir = os.path.join(base_path, template.value.id, "surf", "surf_hybrid_mni_gii")
+    meshes = {}
+    for surf in ["pial", "white"]:
+        for hemi in ["left", "right"]:
+            mesh_path = os.path.join(mesh_dir, f"{hemi[0]}h.{surf}.{target_space}.gii")
+            meshes[f"{surf}_{hemi}"] = load_surf_mesh(mesh_path)
+    return meshes
+
+
+class TribeSurfaceProjector(ns.extractors.neuro.SurfaceProjector):
+    """Project data to an fsaverage surface mesh.
+    For volumetric data, this uses ``nilearn.surface.vol_to_surf`` to project the data to the surface.
+    For surface data, this simply downsamples the data to the target mesh resolution.
+
+    Fields beyond ``mesh`` mirror the keyword arguments of
+    ``nilearn.surface.vol_to_surf`` and are forwarded to it.
+
+    Examples
+    --------
+    >>> SurfaceProjector(mesh="fsaverage5")
+    >>> SurfaceProjector(mesh="fsaverage6", radius=5.0, interpolation="nearest")
+    """
+
+    mesh: str
+    radius: float = 3.0
+    interpolation: tp.Literal["linear", "nearest"] = "linear"
+    kind: tp.Literal["auto", "line", "ball"] = "auto"
+    n_samples: int | None = None
+    mask_img: tp.Any | None = None
+    depth: list[float] | None = None
+    center_depth: float = 1
+    extract_fsaverage_from_mni: bool = False
+
+    _mesh: tp.Any | None = pydantic.PrivateAttr(default=None)
+
+    def model_post_init(self, __context: tp.Any) -> None:
+        super().model_post_init(__context)
+        assert (
+            self.center_depth >= 0 and self.center_depth <= 1
+        ), "center_depth must be between 0 and 1"
+        if self.mesh not in FSAVERAGE_SIZES:
+            raise ValueError(f"mesh must be an fsaverage mesh (got {self.mesh!r})")
+
+    def get_mesh(self) -> tp.Any:
+        if self._mesh is None:
+            if self.extract_fsaverage_from_mni:
+                mni_template_spec = FmriTemplateSpace["MNI152NLIN2009C_ASYM_RES_01"]
+                fsaverage = load_mni_mesh(mni_template_spec, self.mesh)
+            else:
+                from nilearn import datasets
+
+                fsaverage = datasets.fetch_surf_fsaverage(self.mesh)
+            self._mesh = fsaverage
+        return self._mesh
+
+    def get_intermediate_mesh(
+        self, hemi: str, center_depth: float = 0.5
+    ) -> tuple[np.ndarray, np.ndarray]:
+        meshes = self.get_mesh()
+        surf_mesh, inner_mesh = meshes[f"pial_{hemi}"], meshes[f"white_{hemi}"]
+        from nilearn.surface import InMemoryMesh
+
+        if isinstance(surf_mesh, str):
+            import nibabel
+
+            surf_vertices, surf_faces = nibabel.load(surf_mesh).darrays
+            inner_vertices, inner_faces = nibabel.load(inner_mesh).darrays
+            surf_vertices, surf_faces = surf_vertices.data, surf_faces.data
+            inner_vertices, inner_faces = inner_vertices.data, inner_faces.data
+        elif isinstance(surf_mesh, InMemoryMesh):
+            surf_vertices, surf_faces = surf_mesh.coordinates, surf_mesh.faces
+            inner_vertices, inner_faces = inner_mesh.coordinates, inner_mesh.faces
+        else:
+            raise TypeError(f"Unsupported mesh type: {type(surf_mesh)}")
+        half_vertices = surf_vertices * center_depth + inner_vertices * (
+            1 - center_depth
+        )
+        half_depth_mesh = (half_vertices, surf_faces)
+        return half_depth_mesh
+
+    def apply(self, rec: tp.Any) -> np.ndarray:
+
+        if len(rec.shape) == 4:
+            # 4-D volume data → use nilearn.surface.vol_to_surf
+            meshes = self.get_mesh()
+            from nilearn.surface import vol_to_surf
+
+            hemis = []
+            for hemi in ("left", "right"):
+                if self.center_depth == 1:
+                    surf_mesh = meshes[f"pial_{hemi}"]
+                else:
+                    surf_mesh = self.get_intermediate_mesh(hemi, self.center_depth)
+                hemis.append(
+                    vol_to_surf(
+                        rec,
+                        surf_mesh=surf_mesh,
+                        inner_mesh=meshes[f"white_{hemi}"],
+                        radius=self.radius,
+                        interpolation=self.interpolation,
+                        kind=self.kind,
+                        n_samples=self.n_samples,
+                        mask_img=self.mask_img,
+                        depth=self.depth,
+                    )
+                )
+            return np.vstack(hemis)
+
+        elif len(rec.shape) == 2:
+            # 2-D surface data → downsample to target mesh resolution
+            n_vertices = rec.shape[0] // 2
+            if n_vertices not in list(FSAVERAGE_SIZES.values()) or rec.shape[0] % 2:
+                msg = f"The detected number of vertices ({rec.shape[0]}) is not in {list(FSAVERAGE_SIZES.values())}"
+                raise ValueError(msg)
+            n_vertices_resampled = FSAVERAGE_SIZES.get(self.mesh)
+            data = rec.get_fdata()
+            if n_vertices < n_vertices_resampled:
+                raise NotImplementedError(
+                    f"Cannot upsample from {n_vertices} vertices to {n_vertices_resampled} vertices"
+                )
+            if n_vertices > n_vertices_resampled:
+                left = data[:n_vertices_resampled, :]
+                right = data[n_vertices : n_vertices + n_vertices_resampled, :]
+                data = np.concatenate([left, right], axis=0)
+            return data
+        else:
+            raise ValueError(
+                f"Unexpected shape {rec.shape} (should have 2 or 4 dimensions)"
+            )