fix missing t5 model

app.py

@@ -1,11 +1,14 @@
 import streamlit as st
-from transformers import pipeline, PegasusForConditionalGeneration, PegasusTokenizer
+from transformers import pipeline, PegasusForConditionalGeneration, PegasusTokenizer, AutoTokenizer, AutoModelForSeq2SeqLM
 import nltk
 
 from fill_in_summary import FillInSummary
 from paraphrase import PegasusParaphraser
 import question_generator as q
 
+tokenizer = AutoTokenizer.from_pretrained("google/mt5-small")
+
+model = AutoModelForSeq2SeqLM.from_pretrained("google/mt5-small")
 
 # Question Generator Variables
 ids = {'mt5-small': st.secrets['small'],
@@ -25,11 +28,11 @@ if select == "Question Generator":
         #st.selectbox('Model',  ['T5', 'GPT Neo-X'])
 
         # Download all models from drive
-        q.download_models(ids)
+        # q.download_models(ids)
 
         # Model selection
         model_path = st.selectbox('', options=[k for k in ids], index=1, help='Model to use. ')
-        model = q.load_model(model_path=f"model/{model_path}.ckpt")
+
 
         text_input = st.text_area("Input Text")
 
@@ -39,7 +42,7 @@ if select == "Question Generator":
 
         if split:
             # Split into sentences
-            sent_tokenized = nltk.sent_tokenize(inputs)
+            sent_tokenized = nltk.sent_tokenize(text_input)
             res = {}
 
             with st.spinner('Please wait while the inputs are being processed...'):
@@ -61,7 +64,7 @@ if select == "Question Generator":
         else:
             with st.spinner('Please wait while the inputs are being processed...'):
                 # Prediction
-                predictions = model.multitask([inputs], max_length=512)
+                predictions = model.multitask([text_input], max_length=512)
                 questions, answers, answers_bis = predictions['questions'], predictions['answers'], predictions[
                     'answers_bis']
 

question_generator.py

@@ -9,39 +9,6 @@ from transformers import AutoTokenizer
 from mt5 import MT5
 
 
-def download_models(ids):
-    """
-    Download all models.
-    :param ids: name and links of models
-    :return:
-    """
-
-    # Download sentence tokenizer
-    nltk.download('punkt')
-
-    # Download model from drive if not stored locally
-    for key in ids:
-        if not os.path.isfile(f"model/{key}.ckpt"):
-            url = f"https://drive.google.com/u/0/uc?id={ids[key]}"
-            gdown.download(url=url, output=f"model/{key}.ckpt")
-
-
-@st.cache(allow_output_mutation=True)
-def load_model(model_path):
-    """
-    Load model and cache it.
-    :param model_path: path to model
-    :return:
-    """
-
-    device = 'cuda' if torch.cuda.is_available() else 'cpu'
-
-    # Loading model and tokenizer
-    model = MT5.load_from_checkpoint(model_path).eval().to(device)
-    model.tokenizer = AutoTokenizer.from_pretrained('tokenizer')
-
-    return model
-
 # elif task == 'Question Answering':
 #
 #     # Input area