Update app.py

app.py

@@ -1,26 +1,107 @@
 import streamlit as st
-from transformers import pipeline
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, AutoModelForSequenceClassification
+from scipy.special import softmax
+import torch
 
-# Load sentiment-analysis pipeline from Hugging Face
+# Load sentiment model
 @st.cache_resource
-def load_model():
-    return pipeline("sentiment-analysis")
+def load_sentiment_model():
+    model = AutoModelForSequenceClassification.from_pretrained("cardiffnlp/twitter-roberta-base-sentiment")
+    tokenizer = AutoTokenizer.from_pretrained("cardiffnlp/twitter-roberta-base-sentiment")
+    return model, tokenizer
 
-sentiment_pipeline = load_model()
+# Load emotion model
+@st.cache_resource
+def load_emotion_model():
+    model = AutoModelForSequenceClassification.from_pretrained("j-hartmann/emotion-english-distilroberta-base")
+    tokenizer = AutoTokenizer.from_pretrained("j-hartmann/emotion-english-distilroberta-base")
+    return model, tokenizer
+
+# Load T5 paraphrasing model
+@st.cache_resource
+def load_paraphrase_model():
+    model = AutoModelForSeq2SeqLM.from_pretrained("Vamsi/T5_Paraphrase_Paws")
+    tokenizer = AutoTokenizer.from_pretrained("Vamsi/T5_Paraphrase_Paws")
+    return model, tokenizer
+
+# Sentiment analysis
+def get_sentiment(text, model, tokenizer):
+    encoded = tokenizer(text, return_tensors="pt", truncation=True)
+    with torch.no_grad():
+        output = model(**encoded)
+    probs = softmax(output.logits.numpy()[0])
+    labels = ["Negative", "Neutral", "Positive"]
+    return labels[probs.argmax()], float(probs.max()) * 100
+
+# Emotion detection
+def get_emotion(text, model, tokenizer):
+    encoded = tokenizer(text, return_tensors="pt", truncation=True)
+    with torch.no_grad():
+        output = model(**encoded)
+    probs = softmax(output.logits.numpy()[0])
+    labels = ['anger', 'disgust', 'fear', 'joy', 'neutral', 'sadness', 'surprise']
+    return labels[probs.argmax()], float(probs.max()) * 100
 
-# Streamlit UI
-st.title("Smart Sentiment Analyzer 🤖")
-st.markdown("Analyze your sentence using a real AI model (DistilBERT).")
+# Generate feedback
+def generate_feedback(sentiment, emotion):
+    if sentiment == "Negative":
+        if emotion in ["anger", "disgust", "sadness"]:
+            return "⚠️ Your message might sound hurtful or emotionally charged. Consider softening it."
+        else:
+            return "⚠️ Your message may feel negative. Reflect on your tone before sending."
+    elif sentiment == "Neutral":
+        return "😐 Your message seems neutral. That's okay, but clarity and warmth often help."
+    elif sentiment == "Positive":
+        if emotion == "joy":
+            return "😊 Great! Your message feels joyful and likely to be well received."
+        elif emotion == "love":
+            return "💖 Your message feels loving. Expressing emotions like this builds trust."
+        else:
+            return "🙂 Your message is positive, but think about whether it’s being fully understood."
 
-text = st.text_area("Enter your sentence:")
+# Paraphrase / rewrite message
+def rewrite_message(text, model, tokenizer):
+    text = "paraphrase: " + text + " </s>"
+    encoding = tokenizer.encode_plus(text, return_tensors="pt", max_length=128, truncation=True)
+    with torch.no_grad():
+        output = model.generate(
+            input_ids=encoding['input_ids'],
+            attention_mask=encoding['attention_mask'],
+            max_length=128,
+            num_return_sequences=2,
+            num_beams=5,
+            temperature=1.5
+        )
+    rewrites = [tokenizer.decode(o, skip_special_tokens=True) for o in output]
+    return list(set(rewrites))  # remove duplicates
 
-if st.button("Analyze Sentiment"):
+# UI
+st.title("🗣️ Message Tone & Rewrite Checker (Phase 2)")
+st.write("Before you send that message, check how it might be received — and improve it if needed.")
+
+text = st.text_area("✍️ Enter your message here:")
+
+if st.button("Analyze"):
     if text.strip() == "":
-        st.warning("Please enter some text.")
+        st.warning("Please type a message first.")
     else:
-        result = sentiment_pipeline(text)[0]
-        label = result["label"]
-        score = round(result["score"] * 100, 2)
+        sent_model, sent_token = load_sentiment_model()
+        emo_model, emo_token = load_emotion_model()
+
+        sentiment, s_conf = get_sentiment(text, sent_model, sent_token)
+        emotion, e_conf = get_emotion(text, emo_model, emo_token)
+        feedback = generate_feedback(sentiment, emotion)
+
+        st.markdown("### 🧠 Analysis Result")
+        st.write(f"**Sentiment:** {sentiment} ({s_conf:.2f}%)")
+        st.write(f"**Emotion:** {emotion} ({e_conf:.2f}%)")
+        st.markdown("### 💡 Feedback")
+        st.info(feedback)
+
+        st.markdown("---")
+        st.markdown("### ✨ Try Rewriting Your Message")
+        para_model, para_token = load_paraphrase_model()
+        rewrites = rewrite_message(text, para_model, para_token)
 
-        st.success(f"**Sentiment:** {label}")
-        st.info(f"**Confidence:** {score}%")
+        for i, r in enumerate(rewrites, 1):
+            st.write(f"**Version {i}:** {r}")