Upload app.py with huggingface_hub

app.py

@@ -0,0 +1,98 @@
+
+import streamlit as st
+import torch
+from transformers import AutoProcessor, LlavaForConditionalGeneration, BitsAndBytesConfig, CLIPProcessor, CLIPModel
+from peft import PeftModel
+from PIL import Image
+
+st.set_page_config(page_title="Multimodal Risk Engine", page_icon="🛡️", layout="wide")
+
+# --- LOAD MODELS (Smart Caching) ---
+@st.cache_resource
+def load_models():
+    print("🔄 Loading Models...")
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    
+    # A. Stage 1 (CLIP)
+    clip_id = "openai/clip-vit-base-patch32"
+    clip_model = CLIPModel.from_pretrained(clip_id).to(device)
+    clip_processor = CLIPProcessor.from_pretrained(clip_id)
+    
+    # B. Stage 2 (LLaVA)
+    model_id = "llava-hf/llava-1.5-7b-hf"
+    
+    # CPU vs GPU Loading Logic
+    if device == "cuda":
+        bnb_config = BitsAndBytesConfig(
+            load_in_4bit=True, bnb_4bit_use_double_quant=True, 
+            bnb_4bit_quant_type="nf4", bnb_4bit_compute_dtype=torch.float16
+        )
+        base_model = LlavaForConditionalGeneration.from_pretrained(
+            model_id, quantization_config=bnb_config, torch_dtype=torch.float16, device_map="auto"
+        )
+    else:
+        # Fallback for Free CPU Tier (Might be slow/crash but allows build verification)
+        base_model = LlavaForConditionalGeneration.from_pretrained(model_id)
+    
+    adapter_id = "oke39/llava-v1.5-7b-hateful-memes-lora"
+    model = PeftModel.from_pretrained(base_model, adapter_id)
+    llava_processor = AutoProcessor.from_pretrained(model_id)
+    
+    return clip_model, clip_processor, model, llava_processor
+
+try:
+    clip_model, clip_processor, llava_model, llava_processor = load_models()
+    st.toast("✅ System Ready", icon="🚀")
+except Exception as e:
+    st.error(f"Hardware Error: {e}")
+    st.stop()
+
+# --- PIPELINE ---
+def stage_1_glance(image):
+    HATE_ANCHOR = ["hate speech, offensive content, racism, dangerous propaganda"]
+    inputs = clip_processor(text=HATE_ANCHOR, images=image, return_tensors="pt", padding=True).to(clip_model.device)
+    with torch.no_grad():
+        outputs = clip_model(**inputs)
+        probs = outputs.logits_per_image.softmax(dim=1)
+        return float(probs[0][0])
+
+def stage_2_judge(image, text_context):
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    prompt = f"USER: <image>\nAnalyze this meme text: '{text_context}'. Is it hateful? Return JSON.\nASSISTANT:"
+    inputs = llava_processor(text=prompt, images=image, return_tensors="pt").to(device)
+    
+    with torch.inference_mode():
+        output = llava_model.generate(**inputs, max_new_tokens=200)
+    response = llava_processor.decode(output[0], skip_special_tokens=True)
+    return response.split("ASSISTANT:")[-1].strip() if "ASSISTANT:" in response else response
+
+# --- UI ---
+st.title("🛡️ Multimodal Content Risk Engine")
+st.markdown("### The 'Benign Confounder' Solver")
+
+col1, col2 = st.columns([1, 1])
+
+with col1:
+    uploaded_file = st.file_uploader("Upload Meme", type=["jpg", "png", "jpeg"])
+    text_input = st.text_input("Extracted Text", placeholder="Type visible text...")
+    if uploaded_file:
+        image = Image.open(uploaded_file).convert("RGB")
+        st.image(image, caption="User Upload", use_container_width=True)
+
+with col2:
+    if uploaded_file and st.button("Analyze Risk", type="primary"):
+        with st.status("Running Pipeline...", expanded=True) as status:
+            st.write("👁️ **Stage 1: The 'Glance' (CLIP)**")
+            risk_score = stage_1_glance(image)
+            st.progress(min(risk_score, 1.0))
+            
+            if risk_score < 0.22:
+                status.update(label="✅ Auto-Approved", state="complete")
+                st.success("Verdict: BENIGN")
+                st.write(f"Risk Score: `{risk_score:.4f}`")
+            else:
+                st.warning(f"⚠️ Risk Detected ({risk_score:.4f})! Escalating...")
+                st.write("⚖️ **Stage 2: The 'Judge' (LLaVA)**")
+                verdict = stage_2_judge(image, text_input if text_input else "")
+                status.update(label="✅ Done", state="complete")
+                st.json(verdict)