Create app.py

app.py

@@ -0,0 +1,192 @@
+import gradio as gr
+import torch
+import torch.nn as nn
+from transformers import pipeline, BertTokenizer, CLIPProcessor
+from PIL import Image
+import pytesseract
+import cv2
+import numpy as np
+
+# Initialize OCR
+# Note: You need to install tesseract-ocr on your system
+# For Hugging Face Spaces, add: apt-get install -y tesseract-ocr
+# to a file called packages.txt
+
+class MemeAnalyzerWithOCR:
+    def __init__(self):
+        # Sentiment Analysis for text (Positive, Negative, Neutral)
+        self.text_classifier = pipeline(
+            "sentiment-analysis", 
+            model="cardiffnlp/twitter-roberta-base-sentiment-latest"
+        )
+        
+        # Hate Speech Detection for the complete meme
+        self.hate_detector = pipeline(
+            "text-classification",
+            model="unitary/toxic-bert"
+        )
+        
+        # Image understanding (not specifically for hate, but for context)
+        self.image_classifier = pipeline(
+            "image-classification",
+            model="google/vit-base-patch16-224"
+        )
+        
+    def extract_text_from_image(self, image):
+        """Extract text from meme using OCR"""
+        try:
+            # Convert PIL to opencv format
+            image_cv = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
+            
+            # Preprocess image for better OCR
+            gray = cv2.cvtColor(image_cv, cv2.COLOR_BGR2GRAY)
+            # Increase contrast
+            enhanced = cv2.convertScaleAbs(gray, alpha=1.5, beta=0)
+            
+            # Extract text
+            text = pytesseract.image_to_string(enhanced)
+            return text.strip()
+        except Exception as e:
+            return ""
+    
+    def analyze_meme(self, text_input, image):
+        results = {
+            'extracted_text': '',
+            'sentiment': None,
+            'hate_detection': None,
+            'image_content': None,
+            'combined_analysis': ''
+        }
+        
+        # Step 1: Extract text from image if provided
+        if image is not None:
+            extracted_text = self.extract_text_from_image(image)
+            results['extracted_text'] = extracted_text
+            
+            # Analyze image content
+            image_results = self.image_classifier(image)
+            results['image_content'] = image_results[0]['label']
+        
+        # Step 2: Combine manual text input with OCR text
+        combined_text = ""
+        if text_input:
+            combined_text = text_input
+        if results['extracted_text']:
+            combined_text = combined_text + " " + results['extracted_text'] if combined_text else results['extracted_text']
+        
+        if not combined_text:
+            return "No text found! Please provide text or an image with text."
+        
+        # Step 3: Sentiment Analysis (Positive, Negative, Neutral)
+        sentiment_result = self.text_classifier(combined_text)[0]
+        
+        # Map to your categories
+        sentiment_mapping = {
+            'positive': 'Positive',
+            'negative': 'Negative', 
+            'neutral': 'Neutral'
+        }
+        
+        results['sentiment'] = {
+            'label': sentiment_mapping.get(sentiment_result['label'].lower(), 'Neutral'),
+            'score': sentiment_result['score']
+        }
+        
+        # Step 4: Hate Speech Detection
+        hate_result = self.hate_detector(combined_text)[0]
+        
+        # Determine if hateful
+        is_hateful = hate_result['label'] == 'TOXIC' and hate_result['score'] > 0.7
+        results['hate_detection'] = {
+            'label': 'Hateful' if is_hateful else 'Non-hateful',
+            'score': hate_result['score'] if is_hateful else 1 - hate_result['score']
+        }
+        
+        # Step 5: Format results
+        output = "## 📊 Meme Analysis Results\n\n"
+        
+        # Show extracted text
+        if results['extracted_text']:
+            output += f"### 🔍 Text Extracted from Image (OCR):\n`{results['extracted_text']}`\n\n"
+        
+        # Sentiment Analysis
+        output += f"### 😊 Sentiment Analysis (BERT):\n"
+        output += f"**{results['sentiment']['label']}** "
+        output += f"(Confidence: {results['sentiment']['score']:.1%})\n\n"
+        
+        # Hate Detection
+        output += f"### 🚫 Hate Speech Detection:\n"
+        output += f"**{results['hate_detection']['label']}** "
+        output += f"(Confidence: {results['hate_detection']['score']:.1%})\n\n"
+        
+        # Image content
+        if results['image_content']:
+            output += f"### 🖼️ Image Content:\n{results['image_content']}\n\n"
+        
+        # Combined analysis
+        output += "### 📝 Analyzed Text:\n"
+        output += f"`{combined_text}`\n\n"
+        
+        # Warning for hateful content
+        if is_hateful:
+            output += "⚠️ **Warning**: This content may contain hateful or offensive material.\n"
+        
+        return output
+
+# Initialize analyzer
+analyzer = MemeAnalyzerWithOCR()
+
+# Create Gradio interface
+demo = gr.Interface(
+    fn=analyzer.analyze_meme,
+    inputs=[
+        gr.Textbox(
+            label="📝 Manual Text Input (Optional)",
+            placeholder="Enter text if not in image...",
+            info="Leave empty if text is in the image"
+        ),
+        gr.Image(
+            label="📸 Upload Meme Image",
+            type="pil",
+            info="The AI will extract text from the image"
+        )
+    ],
+    outputs=gr.Markdown(label="Analysis Results"),
+    title="🎭 Meme Analyzer with OCR",
+    description="""
+    This tool analyzes memes by:
+    1. **Extracting text** from images using OCR
+    2. **Sentiment analysis** (Positive/Negative/Neutral) using BERT
+    3. **Hate speech detection** (Hateful/Non-hateful)
+    4. **Image content analysis**
+    
+    Upload a meme image and/or provide text to analyze!
+    """,
+    examples=[
+        ["", "examples/meme1.jpg"],
+        ["This is hilarious!", None],
+        ["I hate everyone", None]
+    ],
+    theme=gr.themes.Soft()
+)
+
+# Launch
+demo.launch()
+
+# For Hugging Face Spaces, create these additional files:
+
+# requirements.txt:
+"""
+gradio
+torch
+transformers
+pillow
+opencv-python
+pytesseract
+numpy
+"""
+
+# packages.txt (for system dependencies):
+"""
+tesseract-ocr
+"""