Switched to FLAN-T5-Small and added UX improvements

app.py

@@ -1,36 +1,36 @@
 import streamlit as st
-import torch
 import time
-from transformers import T5Tokenizer, T5ForConditionalGeneration
+import os 
+from transformers import T5Tokenizer, TFT5ForConditionalGeneration # NEW: Using TF variant
+
+# --- FIX 1: Set a safe cache directory to avoid PermissionError at /.cache ---
+os.environ["HF_HOME"] = "/tmp/hf_cache" 
 
 # --- Configuration ---
-# Using FLAN-T5-Small model (approx. 80MB) for fast startup
 MODEL_NAME = "google/flan-t5-small" 
 
 # -------------------- Model Logic --------------------
 
-# Define the system prompt with tuning incorporated
 sys_prompt = """
 Categorize the sentiment of the customer review as positive, negative, or neutral.
-Leverage your expertise in the aviation industry and deep understanding of industry trends to analyze the nuanced expressions and overall tone.
-It is crucial to accurately identify neutral sentiments, which may indicate a balanced view or neutral stance towards the airline. Neutral expressions could involve factual statements without explicit positive or negative opinions.
-Consider the importance of these neutral sentiments in gauging the public sentiment towards the airline company.
-For instance, a positive sentiment might convey satisfaction with the airline's services, a negative sentiment could express dissatisfaction, while neutral sentiment may reflect an impartial observation or a neutral standpoint
+Respond ONLY with the sentiment word (positive, negative, or neutral). Do NOT include any other text.
 """
 
 @st.cache_resource
 def load_llm():
-    """Loads the model and tokenizer with a friendly loading message."""
+    """Loads the model and tokenizer without PyTorch."""
+    # Since we removed torch, device is always 'cpu' conceptually
+    device = "CPU (TensorFlow)"
+    
     try:
-        with st.spinner(f"Loading tokenizer and model ({MODEL_NAME}). This is the slowest part!"):
-            # Determine the device
-            device = "cuda" if torch.cuda.is_available() else "cpu"
-            st.info(f"Using device: **{device.upper()}**. Starting model download...")
+        with st.spinner(f"Loading tokenizer and model ({MODEL_NAME}) on {device}..."):
+            st.info(f"Using device: **{device}**. Starting model download...")
             
             start_time = time.time()
             tokenizer = T5Tokenizer.from_pretrained(MODEL_NAME)
-            # We explicitly place the model on the determined device
-            model = T5ForConditionalGeneration.from_pretrained(MODEL_NAME).to(device)
+            
+            # CRUCIAL CHANGE: Load the TensorFlow version of the model
+            model = TFT5ForConditionalGeneration.from_pretrained(MODEL_NAME)
             
             end_time = time.time()
             st.success(f"Model {MODEL_NAME} loaded successfully in {end_time - start_time:.2f} seconds!")
@@ -39,15 +39,16 @@ def load_llm():
             
     except Exception as e:
         st.error(f"FATAL ERROR LOADING MODEL: {e}")
-        st.info("Check the Space logs for dependency or memory errors.")
+        st.info("Model load failed. Check the logs for missing TensorFlow dependencies.")
         return None, None, None
 
 def llm_response(tokenizer, model, device, prompt):
     if tokenizer is None or model is None:
         return "Model not initialized due to previous error."
     
-    # Process and move input to the correct device
-    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
+    # Process input (TensorFlow models automatically handle device placement on CPU)
+    # TF models use the 'tf' argument for return tensors
+    input_ids = tokenizer(prompt, return_tensors="tf").input_ids
     
     # Generate response
     outputs = model.generate(input_ids, max_length=150, do_sample=False)
@@ -56,12 +57,7 @@ def llm_response(tokenizer, model, device, prompt):
     return tokenizer.decode(outputs[0], skip_special_tokens=True).strip()
 
 def predict_review_sentiment(tokenizer, model, device, review):
-    """Function to format the prompt and get the prediction."""
-    full_prompt = """
-            {}
-            Review text: '{}'
-    """.format(sys_prompt, review)
-    
+    full_prompt = f"{sys_prompt}\nReview text: '{review}'"
     return llm_response(tokenizer, model, device, full_prompt)
 
 
@@ -70,42 +66,46 @@ def predict_review_sentiment(tokenizer, model, device, review):
 # --- 1. Load Resources ---
 tokenizer, model, device = load_llm()
 
-# --- 2. System Info Message (Simplified) ---
-st.sidebar.markdown("## 📊 System Info (Check Space Logs for Live Monitor)")
-st.sidebar.info(
-    f"Model is: **{MODEL_NAME}**\n\n"
-    f"Device: **{device.upper()}**\n\n"
-    "Status: **App Fully Loaded**"
-)
-
-# --- 3. Main Application ---
-st.title("✈️ Customer Review Sentiment Analyzer")
-st.markdown(f"Using the ultra-lightweight **{MODEL_NAME}** model for fast sentiment classification.")
-
-st.subheader("Enter the Customer Review:")
-
-# Collect user input
-review_text = st.text_area(
-    "Customer Review:", 
-    height=150, 
-    placeholder="E.g., I'm stuck on the tarmac for two hours. This is unacceptable! #SkyWingsfail"
-)
-
-# Predict button
-if st.button("Predict Sentiment"):
-    if not review_text:
-        st.warning("Please enter a customer review to predict the sentiment.")
-    elif tokenizer is not None:
-        with st.spinner('Analyzing sentiment...'):
-            pred_sent = predict_review_sentiment(tokenizer, model, device, review_text)
-            
-            # Display result with color coding
-            sentiment_text = pred_sent.lower()
-            if 'positive' in sentiment_text:
-                st.success(f"**Predicted Sentiment:** {pred_sent}")
-            elif 'negative' in sentiment_text:
-                st.error(f"**Predicted Sentiment:** {pred_sent}")
-            elif 'neutral' in sentiment_text:
-                st.info(f"**Predicted Sentiment:** {pred_sent}")
-            else:
-                st.warning(f"**Predicted Sentiment (Unknown Category):** {pred_sent}")
+# --- 2. System Info Message (Conditional Display) ---
+st.sidebar.markdown("## 📊 Deployment Status")
+
+if tokenizer is not None:
+    st.sidebar.info(
+        f"Model: **{MODEL_NAME}**\n\n"
+        f"Framework: **{device}**\n\n"
+        "Status: **Local Inference Ready**"
+    )
+
+    # --- 3. Main Application ---
+    st.title("✈️ Customer Review Sentiment Analyzer")
+    st.markdown(f"Using the TensorFlow-backed **{MODEL_NAME}** model.")
+
+    st.subheader("Enter the Customer Review:")
+
+    review_text = st.text_area(
+        "Customer Review:", 
+        height=150, 
+        placeholder="E.g., The flight was delayed, but the crew was excellent."
+    )
+
+    # Predict button
+    if st.button("Predict Sentiment"):
+        if not review_text:
+            st.warning("Please enter a customer review to predict the sentiment.")
+        else:
+            with st.spinner('Analyzing sentiment...'):
+                pred_sent = predict_review_sentiment(tokenizer, model, device, review_text)
+                
+                # Display result with color coding
+                sentiment_text = pred_sent.lower()
+                if 'positive' in sentiment_text:
+                    st.success(f"**Predicted Sentiment:** {pred_sent}")
+                elif 'negative' in sentiment_text:
+                    st.error(f"**Predicted Sentiment:** {pred_sent}")
+                elif 'neutral' in sentiment_text:
+                    st.info(f"**Predicted Sentiment:** {pred_sent}")
+                else:
+                    st.warning(f"**Predicted Sentiment (Model Response):** {pred_sent}")
+else:
+    # Display a clear error on the main page if the model load failed
+    st.error("Application Failed to Initialize. Model load aborted (likely missing TensorFlow). Check Space logs.")