Update app.py

app.py

@@ -11,6 +11,17 @@ from datetime import datetime
 from torch.nn.functional import sigmoid
 from collections import Counter
 import logging
+from transformers import pipeline as hf_pipeline
+
+# Add this with your other model loading code
+emotion_pipeline = hf_pipeline(
+    "text-classification",
+    model="j-hartmann/emotion-english-distilroberta-base",
+    top_k=6,
+    truncation=True,
+    device=0 if torch.cuda.is_available() else -1  # GPU support
+)
+
 # Add this after imports
 device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
 
@@ -26,9 +37,6 @@ model_name = "SamanthaStorm/tether-multilabel-v4"
 model = AutoModelForSequenceClassification.from_pretrained(model_name).to(device)
 tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
 
-# Tone model
-tone_model = AutoModelForSequenceClassification.from_pretrained("SamanthaStorm/tone-tag-multilabel-v1").to(device)
-tone_tokenizer = AutoTokenizer.from_pretrained("SamanthaStorm/tone-tag-multilabel-v1", use_fast=False)
 
 # Sentiment model
 sentiment_model = AutoModelForSequenceClassification.from_pretrained("SamanthaStorm/tether-sentiment").to(device)
@@ -126,6 +134,142 @@ THREAT_MOTIFS = [
     "if you just behaved, this wouldn't happen", "this is your fault",
     "you're making me hurt you", "i warned you", "you should have listened"
 ]
+def get_emotion_profile(text):
+    emotions = emotion_pipeline(text)
+    if isinstance(emotions, list) and isinstance(emotions[0], list):
+        emotions = emotions[0]
+    return {e['label'].lower(): round(e['score'], 3) for e in emotions}
+
+def get_emotional_tone_tag(text, sentiment, patterns, abuse_score):
+    # Get emotions first
+    emotions = get_emotion_profile(text)
+    
+    sadness = emotions.get("sadness", 0)
+    joy = emotions.get("joy", 0)
+    neutral = emotions.get("neutral", 0)
+    disgust = emotions.get("disgust", 0)
+    anger = emotions.get("anger", 0)
+    fear = emotions.get("fear", 0)
+
+      # 1. Performative Regret
+    if (
+        sadness > 0.4 and
+        any(p in patterns for p in ["blame shifting", "guilt tripping", "recovery phase"]) and
+        (sentiment == "undermining" or abuse_score > 40)
+    ):
+        return "performative regret"
+
+    # 2. Coercive Warmth
+    if (
+        (joy > 0.3 or sadness > 0.4) and
+        any(p in patterns for p in ["control", "gaslighting"]) and
+        sentiment == "undermining"
+    ):
+        return "coercive warmth"
+
+    # 3. Cold Invalidation
+    if (
+        (neutral + disgust) > 0.5 and
+        any(p in patterns for p in ["dismissiveness", "projection", "obscure language"]) and
+        sentiment == "undermining"
+    ):return "cold invalidation"
+
+    # 4. Genuine Vulnerability
+    if (
+        (sadness + fear) > 0.5 and
+        sentiment == "supportive" and
+        all(p in ["recovery phase"] for p in patterns)
+    ):
+        return "genuine vulnerability"
+
+    # 5. Emotional Threat
+    if (
+        (anger + disgust) > 0.5 and
+        any(p in patterns for p in ["control", "insults", "dismissiveness"]) and
+        sentiment == "undermining"
+    ):
+        return "emotional threat"
+
+    # 6. Weaponized Sadness
+    if (
+        sadness > 0.6 and
+        any(p in patterns for p in ["guilt tripping", "projection"]) and
+        sentiment == "undermining"
+    ):
+        return "weaponized sadness"
+
+    # 7. Toxic Resignation
+    if (
+        neutral > 0.5 and
+        any(p in patterns for p in ["dismissiveness", "obscure language"]) and
+        sentiment == "undermining"
+    ):
+        return "toxic resignation"
+     # 8. Aggressive Dismissal
+    if (
+        anger > 0.5 and
+        any(p in patterns for p in ["aggression", "insults", "control"]) and
+        sentiment == "undermining"
+    ):
+        return "aggressive dismissal"
+    # 9. Deflective Hostility
+    if (
+        (0.2 < anger < 0.7 or 0.2 < disgust < 0.7) and
+        any(p in patterns for p in ["deflection", "projection"]) and
+        sentiment == "undermining"
+    ):
+        return "deflective hostility"   
+    # 10. Mocking Detachment
+    if (
+        (neutral + joy) > 0.5 and
+        any(p in patterns for p in ["mockery", "insults", "projection"]) and
+        sentiment == "undermining"
+    ):
+        return "mocking detachment"
+        # 11. Contradictory Gaslight
+    if (
+        (joy + anger + sadness) > 0.5 and
+        any(p in patterns for p in ["gaslighting", "contradictory statements"]) and
+        sentiment == "undermining"
+    ):
+        return "contradictory gaslight"
+        # 12. Calculated Neutrality
+    if (
+        neutral > 0.6 and
+        any(p in patterns for p in ["obscure language", "deflection", "dismissiveness"]) and
+        sentiment == "undermining"
+    ):
+        return "calculated neutrality"
+     # 13. Forced Accountability Flip
+    if (
+        (anger + disgust) > 0.5 and
+        any(p in patterns for p in ["blame shifting", "manipulation", "projection"]) and
+        sentiment == "undermining"
+    ):
+        return "forced accountability flip"
+        # 14. Conditional Affection
+    if (
+        joy > 0.4 and
+        any(p in patterns for p in ["apology baiting", "control", "recovery phase"]) and
+        sentiment == "undermining"
+    ):
+        return "conditional affection"
+    
+    if (
+        (anger + disgust) > 0.5 and
+        any(p in patterns for p in ["blame shifting", "projection", "deflection"]) and
+        sentiment == "undermining"
+    ):
+        return "forced accountability flip"
+
+    # Emotional Instability Fallback
+    if (
+        (anger + sadness + disgust) > 0.6 and
+        sentiment == "undermining"
+    ):
+        return "emotional instability"
+        
+    return None
 
 def predict_darvo_score(text):
     """Predict DARVO score for given text"""
@@ -161,19 +305,6 @@ def get_risk_stage(patterns, sentiment):
         logger.error(f"Error determining risk stage: {e}")
         return 1
 
-def get_emotional_tone_tag(text, emotions, sentiment, patterns, abuse_score):
-    """Get emotional tone tag for text"""
-    try:
-        inputs = tone_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
-        inputs = {k: v.to(device) for k, v in inputs.items()}
-        with torch.no_grad():
-            logits = tone_model(**inputs).logits[0]
-        probs = torch.sigmoid(logits).cpu().numpy()
-        scores = dict(zip(TONE_LABELS, np.round(probs, 3)))
-        return max(scores, key=scores.get)
-    except Exception as e:
-        logger.error(f"Error in emotional tone analysis: {e}")
-        return "unknown"
 
 @spaces.GPU
 def compute_abuse_score(matched_scores, sentiment):
@@ -243,7 +374,16 @@ def analyze_single_message(text, thresholds):
         explicit_abuse_words = ['fuck', 'bitch', 'shit', 'ass', 'dick']
         explicit_abuse = any(word in text.lower() for word in explicit_abuse_words)
         logger.debug(f"Explicit abuse detected: {explicit_abuse}")
+         # Get sentiment
+        sent_inputs = sentiment_tokenizer(text, return_tensors="pt", truncation=True, padding=True)
+        sent_inputs = {k: v.to(device) for k, v in sent_inputs.items()}
+        with torch.no_grad():
+            sent_logits = sentiment_model(**sent_inputs).logits[0]
+        sent_probs = torch.softmax(sent_logits, dim=-1).cpu().numpy()
+        sentiment = SENTIMENT_LABELS[int(np.argmax(sent_probs))]
         
+        # Get tone using emotion-based approach
+        tone_tag = get_emotional_tone_tag(text, sentiment, threshold_labels, abuse_score)
         # Abuse model inference 
         inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
         inputs = {k: v.to(device) for k, v in inputs.items()}  # Move to GPU