Update app.py

app.py

@@ -65,13 +65,12 @@ app = gr.mount_gradio_app(app, gui, path="/")
 @app.get("/")
 def home():
     return RedirectResponse(url="/") """
-from fastapi import FastAPI
-from fastapi.responses import RedirectResponse
+from fastapi import FastAPI, UploadFile, Form
+from fastapi.responses import RedirectResponse, FileResponse, JSONResponse
+import os
+import shutil
 from PIL import Image
-from transformers import (
-    ViltProcessor, ViltForQuestionAnswering,
-    T5Tokenizer, T5ForConditionalGeneration
-)
+from transformers import ViltProcessor, ViltForQuestionAnswering, AutoTokenizer, AutoModelForCausalLM
 from gtts import gTTS
 import torch
 import tempfile
@@ -79,20 +78,29 @@ import gradio as gr
 
 app = FastAPI()
 
-# VQA Model
+# Load VQA Model
 vqa_processor = ViltProcessor.from_pretrained("dandelin/vilt-b32-finetuned-vqa")
 vqa_model = ViltForQuestionAnswering.from_pretrained("dandelin/vilt-b32-finetuned-vqa")
 
-# Text Rewriter (FLAN-T5-base)
-rewrite_tokenizer = T5Tokenizer.from_pretrained("google/flan-t5-base")
-rewrite_model = T5ForConditionalGeneration.from_pretrained("google/flan-t5-base")
+# Load GPT model to rewrite answers
+# Replacing Falcon-7B-Instruct with GPT-Neo-125M
+gpt_tokenizer = AutoTokenizer.from_pretrained("EleutherAI/gpt-neo-125M")
+gpt_model = AutoModelForCausalLM.from_pretrained("EleutherAI/gpt-neo-125M")
 
 def rewrite_answer(question, short_answer):
-    prompt = f"Answer the question '{question}' with a complete sentence using this answer: '{short_answer}'"
-    inputs = rewrite_tokenizer(prompt, return_tensors="pt")
+    prompt = f"Question: {question}\nShort Answer: {short_answer}\nFull Sentence:"  # Simpler prompt for GPT-Neo
+    inputs = gpt_tokenizer(prompt, return_tensors="pt")
     with torch.no_grad():
-        outputs = rewrite_model.generate(**inputs, max_new_tokens=50)
-    return rewrite_tokenizer.decode(outputs[0], skip_special_tokens=True)
+        outputs = gpt_model.generate(
+            **inputs,
+            max_new_tokens=50,
+            do_sample=True,
+            top_p=0.9,
+            temperature=0.8,
+            pad_token_id=gpt_tokenizer.eos_token_id
+        )
+    full = gpt_tokenizer.decode(outputs[0], skip_special_tokens=True)
+    return full.split("Full Sentence:")[-1].strip()
 
 def answer_question_from_image(image, question):
     if image is None or not question.strip():
@@ -104,21 +112,23 @@ def answer_question_from_image(image, question):
     predicted_id = outputs.logits.argmax(-1).item()
     short_answer = vqa_model.config.id2label[predicted_id]
 
-    # Rewrite to full sentence
+    # Rewrite short answer to full sentence
     full_answer = rewrite_answer(question, short_answer)
 
     try:
         tts = gTTS(text=full_answer)
         with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as tmp:
             tts.save(tmp.name)
-            return full_answer, tmp.name
+            audio_path = tmp.name
     except Exception as e:
-        return f"{full_answer}\n\n⚠️ Audio generation error: {e}", None
+        return f"Answer: {full_answer}\n\n⚠️ Audio generation error: {e}", None
+
+    return full_answer, audio_path
 
 def process_image_question(image: Image.Image, question: str):
-    return answer_question_from_image(image, question)
+    answer, audio_path = answer_question_from_image(image, question)
+    return answer, audio_path
 
-# Gradio UI
 gui = gr.Interface(
     fn=process_image_question,
     inputs=[