Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,663 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# app.py - Enhanced Ensemble Model for Meme and Text Analysis
|
2 |
+
import gradio as gr
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import numpy as np
|
6 |
+
from PIL import Image
|
7 |
+
import requests
|
8 |
+
from io import BytesIO
|
9 |
+
import easyocr
|
10 |
+
import cv2
|
11 |
+
import re
|
12 |
+
from urllib.parse import urlparse
|
13 |
+
import json
|
14 |
+
import logging
|
15 |
+
from typing import Dict, List, Tuple, Optional
|
16 |
+
import warnings
|
17 |
+
warnings.filterwarnings("ignore")
|
18 |
+
|
19 |
+
# Set up logging
|
20 |
+
logging.basicConfig(level=logging.INFO)
|
21 |
+
logger = logging.getLogger(__name__)
|
22 |
+
|
23 |
+
# Import transformers components
|
24 |
+
from transformers import (
|
25 |
+
AutoTokenizer, AutoModelForSequenceClassification,
|
26 |
+
AutoProcessor, AutoModel, SiglipVisionModel,
|
27 |
+
SiglipProcessor, pipeline
|
28 |
+
)
|
29 |
+
|
30 |
+
class EnhancedEnsembleMemeAnalyzer:
|
31 |
+
def __init__(self):
|
32 |
+
"""Initialize the enhanced ensemble model with best available models"""
|
33 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
34 |
+
logger.info(f"Using device: {self.device}")
|
35 |
+
|
36 |
+
# Initialize models
|
37 |
+
self.setup_models()
|
38 |
+
self.setup_ocr()
|
39 |
+
self.setup_ensemble_weights()
|
40 |
+
|
41 |
+
def setup_models(self):
|
42 |
+
"""Initialize BERT and SigLIP models with error handling"""
|
43 |
+
try:
|
44 |
+
# Load your fine-tuned BERT model (93% accuracy)
|
45 |
+
logger.info("Loading fine-tuned BERT model...")
|
46 |
+
self.bert_tokenizer = AutoTokenizer.from_pretrained("./fine_tuned_bert_sentiment")
|
47 |
+
self.bert_model = AutoModelForSequenceClassification.from_pretrained("./fine_tuned_bert_sentiment")
|
48 |
+
self.bert_model.to(self.device)
|
49 |
+
logger.info("β
Fine-tuned BERT loaded successfully!")
|
50 |
+
|
51 |
+
except Exception as e:
|
52 |
+
logger.warning(f"β οΈ Could not load custom BERT model: {e}")
|
53 |
+
logger.info("Loading fallback BERT model...")
|
54 |
+
# Fallback to high-performance public model
|
55 |
+
self.bert_tokenizer = AutoTokenizer.from_pretrained("cardiffnlp/twitter-roberta-base-sentiment-latest")
|
56 |
+
self.bert_model = AutoModelForSequenceClassification.from_pretrained("cardiffnlp/twitter-roberta-base-sentiment-latest")
|
57 |
+
self.bert_model.to(self.device)
|
58 |
+
|
59 |
+
try:
|
60 |
+
# Load the best available SigLIP model (Large version)
|
61 |
+
logger.info("Loading SigLIP-Large model...")
|
62 |
+
self.siglip_processor = AutoProcessor.from_pretrained("google/siglip-large-patch16-384")
|
63 |
+
self.siglip_model = AutoModel.from_pretrained("google/siglip-large-patch16-384")
|
64 |
+
self.siglip_model.to(self.device)
|
65 |
+
|
66 |
+
# Enhanced hate speech classifier on top of SigLIP features
|
67 |
+
self.hate_classifier = nn.Sequential(
|
68 |
+
nn.Linear(1152, 512), # SigLIP-Large has 1152 dim features
|
69 |
+
nn.ReLU(),
|
70 |
+
nn.Dropout(0.3),
|
71 |
+
nn.Linear(512, 256),
|
72 |
+
nn.ReLU(),
|
73 |
+
nn.Dropout(0.2),
|
74 |
+
nn.Linear(256, 4) # Multi-class: safe, hateful, offensive, spam
|
75 |
+
).to(self.device)
|
76 |
+
|
77 |
+
logger.info("β
SigLIP-Large loaded successfully!")
|
78 |
+
|
79 |
+
except Exception as e:
|
80 |
+
logger.warning(f"β οΈ Could not load SigLIP-Large, trying base model: {e}")
|
81 |
+
# Fallback to base model
|
82 |
+
self.siglip_processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")
|
83 |
+
self.siglip_model = AutoModel.from_pretrained("google/siglip-base-patch16-224")
|
84 |
+
self.siglip_model.to(self.device)
|
85 |
+
|
86 |
+
self.hate_classifier = nn.Sequential(
|
87 |
+
nn.Linear(768, 256),
|
88 |
+
nn.ReLU(),
|
89 |
+
nn.Dropout(0.2),
|
90 |
+
nn.Linear(256, 4)
|
91 |
+
).to(self.device)
|
92 |
+
|
93 |
+
def setup_ocr(self):
|
94 |
+
"""Initialize OCR with multiple engines for better accuracy"""
|
95 |
+
try:
|
96 |
+
# Primary OCR: EasyOCR (good for memes)
|
97 |
+
self.ocr_reader = easyocr.Reader(['en'], gpu=torch.cuda.is_available())
|
98 |
+
logger.info("β
EasyOCR initialized")
|
99 |
+
|
100 |
+
# Backup OCR: We'll use cv2 + basic text detection as fallback
|
101 |
+
self.use_easyocr = True
|
102 |
+
|
103 |
+
except Exception as e:
|
104 |
+
logger.warning(f"β οΈ OCR initialization issue: {e}")
|
105 |
+
self.use_easyocr = False
|
106 |
+
|
107 |
+
def setup_ensemble_weights(self):
|
108 |
+
"""Initialize ensemble weights and thresholds"""
|
109 |
+
self.ensemble_weights = {
|
110 |
+
'text_sentiment': 0.4,
|
111 |
+
'image_content': 0.35,
|
112 |
+
'multimodal_context': 0.25
|
113 |
+
}
|
114 |
+
|
115 |
+
self.risk_thresholds = {
|
116 |
+
'high_risk': 0.8,
|
117 |
+
'medium_risk': 0.6,
|
118 |
+
'low_risk': 0.4
|
119 |
+
}
|
120 |
+
|
121 |
+
# Hate speech keywords for additional context
|
122 |
+
self.hate_keywords = [
|
123 |
+
'hate', 'kill', 'death', 'violence', 'attack',
|
124 |
+
'discriminate', 'racist', 'nazi', 'terrorist'
|
125 |
+
]
|
126 |
+
|
127 |
+
def extract_text_from_image(self, image: Image.Image) -> str:
|
128 |
+
"""Enhanced OCR text extraction with multiple methods"""
|
129 |
+
extracted_texts = []
|
130 |
+
|
131 |
+
try:
|
132 |
+
if self.use_easyocr:
|
133 |
+
# Method 1: EasyOCR
|
134 |
+
img_array = np.array(image)
|
135 |
+
results = self.ocr_reader.readtext(img_array, detail=0)
|
136 |
+
if results:
|
137 |
+
easyocr_text = ' '.join(results)
|
138 |
+
extracted_texts.append(easyocr_text)
|
139 |
+
logger.info(f"EasyOCR extracted: {easyocr_text[:100]}...")
|
140 |
+
|
141 |
+
# Method 2: Basic OpenCV preprocessing + simple text detection
|
142 |
+
img_cv = cv2.cvtColor(np.array(image), cv2.COLOR_RGB2BGR)
|
143 |
+
gray = cv2.cvtColor(img_cv, cv2.COLOR_BGR2GRAY)
|
144 |
+
|
145 |
+
# Enhance text regions
|
146 |
+
kernel = np.ones((1,1), np.uint8)
|
147 |
+
processed = cv2.morphologyEx(gray, cv2.MORPH_CLOSE, kernel)
|
148 |
+
|
149 |
+
# This is a simplified approach - in production you'd use more sophisticated methods
|
150 |
+
|
151 |
+
except Exception as e:
|
152 |
+
logger.error(f"OCR Error: {e}")
|
153 |
+
|
154 |
+
# Combine and clean extracted text
|
155 |
+
final_text = ' '.join(extracted_texts) if extracted_texts else ""
|
156 |
+
return self.clean_text(final_text)
|
157 |
+
|
158 |
+
def clean_text(self, text: str) -> str:
|
159 |
+
"""Clean and preprocess text"""
|
160 |
+
if not text:
|
161 |
+
return ""
|
162 |
+
|
163 |
+
# Remove extra whitespace and special characters
|
164 |
+
text = re.sub(r'\s+', ' ', text)
|
165 |
+
text = re.sub(r'[^\w\s\.\!\?\,\-\:\;\(\)]', '', text)
|
166 |
+
|
167 |
+
return text.strip().lower()
|
168 |
+
|
169 |
+
def analyze_sentiment(self, text: str) -> Dict:
|
170 |
+
"""Analyze sentiment using fine-tuned BERT with confidence calibration"""
|
171 |
+
if not text.strip():
|
172 |
+
return {"label": "NEUTRAL", "score": 0.5, "probabilities": [0.33, 0.34, 0.33]}
|
173 |
+
|
174 |
+
try:
|
175 |
+
inputs = self.bert_tokenizer(
|
176 |
+
text,
|
177 |
+
return_tensors="pt",
|
178 |
+
truncation=True,
|
179 |
+
padding=True,
|
180 |
+
max_length=512
|
181 |
+
).to(self.device)
|
182 |
+
|
183 |
+
with torch.no_grad():
|
184 |
+
outputs = self.bert_model(**inputs)
|
185 |
+
probabilities = torch.nn.functional.softmax(outputs.logits, dim=-1)
|
186 |
+
|
187 |
+
# Get predictions
|
188 |
+
predicted_class = torch.argmax(probabilities, dim=-1).item()
|
189 |
+
confidence = torch.max(probabilities).item()
|
190 |
+
probs_list = probabilities[0].cpu().tolist()
|
191 |
+
|
192 |
+
# Map to sentiment labels (adjust based on your model's configuration)
|
193 |
+
if len(probs_list) == 3:
|
194 |
+
label_mapping = {0: "NEGATIVE", 1: "NEUTRAL", 2: "POSITIVE"}
|
195 |
+
else:
|
196 |
+
label_mapping = {0: "NEGATIVE", 1: "POSITIVE"}
|
197 |
+
|
198 |
+
return {
|
199 |
+
"label": label_mapping.get(predicted_class, "UNKNOWN"),
|
200 |
+
"score": confidence,
|
201 |
+
"probabilities": probs_list
|
202 |
+
}
|
203 |
+
|
204 |
+
except Exception as e:
|
205 |
+
logger.error(f"Sentiment analysis error: {e}")
|
206 |
+
return {"label": "NEUTRAL", "score": 0.5, "probabilities": [0.5, 0.5]}
|
207 |
+
|
208 |
+
def classify_multimodal_content(self, image: Image.Image, text: str = "") -> Dict:
|
209 |
+
"""Enhanced multimodal classification using SigLIP"""
|
210 |
+
try:
|
211 |
+
# Prepare comprehensive text queries for zero-shot classification
|
212 |
+
hate_queries = [
|
213 |
+
"hateful meme targeting specific groups",
|
214 |
+
"discriminatory content with offensive imagery",
|
215 |
+
"violent or threatening visual content",
|
216 |
+
"meme promoting hatred or discrimination",
|
217 |
+
"offensive visual propaganda",
|
218 |
+
"cyberbullying visual content"
|
219 |
+
]
|
220 |
+
|
221 |
+
safe_queries = [
|
222 |
+
"harmless funny meme",
|
223 |
+
"positive social media content",
|
224 |
+
"safe entertainment image",
|
225 |
+
"normal social media post",
|
226 |
+
"friendly humorous content",
|
227 |
+
"non-offensive visual content"
|
228 |
+
]
|
229 |
+
|
230 |
+
# Include context from extracted text
|
231 |
+
if text:
|
232 |
+
context_query = f"image with text saying: {text[:100]}"
|
233 |
+
hate_queries.append(f"hateful {context_query}")
|
234 |
+
safe_queries.append(f"harmless {context_query}")
|
235 |
+
|
236 |
+
all_queries = hate_queries + safe_queries
|
237 |
+
|
238 |
+
# Process with SigLIP
|
239 |
+
inputs = self.siglip_processor(
|
240 |
+
text=all_queries,
|
241 |
+
images=image,
|
242 |
+
return_tensors="pt",
|
243 |
+
padding=True
|
244 |
+
).to(self.device)
|
245 |
+
|
246 |
+
with torch.no_grad():
|
247 |
+
outputs = self.siglip_model(**inputs)
|
248 |
+
logits_per_image = outputs.logits_per_image
|
249 |
+
probs = torch.softmax(logits_per_image, dim=-1)
|
250 |
+
|
251 |
+
# Calculate hate vs safe probabilities
|
252 |
+
hate_prob = torch.sum(probs[0][:len(hate_queries)]).item()
|
253 |
+
safe_prob = torch.sum(probs[0][len(hate_queries):]).item()
|
254 |
+
|
255 |
+
# Normalize probabilities
|
256 |
+
total_prob = hate_prob + safe_prob
|
257 |
+
if total_prob > 0:
|
258 |
+
hate_prob /= total_prob
|
259 |
+
safe_prob /= total_prob
|
260 |
+
|
261 |
+
# Additional keyword-based adjustment
|
262 |
+
keyword_boost = self.check_hate_keywords(text)
|
263 |
+
hate_prob = min(1.0, hate_prob + keyword_boost * 0.1)
|
264 |
+
|
265 |
+
return {
|
266 |
+
"is_hateful": hate_prob > 0.5,
|
267 |
+
"hate_probability": hate_prob,
|
268 |
+
"safe_probability": safe_prob,
|
269 |
+
"confidence": abs(hate_prob - 0.5) * 2,
|
270 |
+
"detailed_scores": probs[0].cpu().tolist()
|
271 |
+
}
|
272 |
+
|
273 |
+
except Exception as e:
|
274 |
+
logger.error(f"Multimodal classification error: {e}")
|
275 |
+
return {
|
276 |
+
"is_hateful": False,
|
277 |
+
"hate_probability": 0.3,
|
278 |
+
"safe_probability": 0.7,
|
279 |
+
"confidence": 0.5,
|
280 |
+
"detailed_scores": []
|
281 |
+
}
|
282 |
+
|
283 |
+
def check_hate_keywords(self, text: str) -> float:
|
284 |
+
"""Check for hate speech keywords and return boost factor"""
|
285 |
+
if not text:
|
286 |
+
return 0.0
|
287 |
+
|
288 |
+
text_lower = text.lower()
|
289 |
+
keyword_count = sum(1 for keyword in self.hate_keywords if keyword in text_lower)
|
290 |
+
|
291 |
+
return min(1.0, keyword_count * 0.2) # Cap at 1.0
|
292 |
+
|
293 |
+
def fetch_social_media_content(self, url: str) -> Dict:
|
294 |
+
"""Enhanced social media content fetching with better error handling"""
|
295 |
+
try:
|
296 |
+
headers = {
|
297 |
+
'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/91.0.4472.124 Safari/537.36'
|
298 |
+
}
|
299 |
+
|
300 |
+
response = requests.get(url, headers=headers, timeout=15)
|
301 |
+
response.raise_for_status()
|
302 |
+
|
303 |
+
content_type = response.headers.get('content-type', '').lower()
|
304 |
+
|
305 |
+
# Handle direct image URLs
|
306 |
+
if any(img_type in content_type for img_type in ['image/jpeg', 'image/png', 'image/gif', 'image/webp']):
|
307 |
+
image = Image.open(BytesIO(response.content))
|
308 |
+
return {"type": "image", "content": image, "url": url}
|
309 |
+
|
310 |
+
# Handle HTML content (simplified scraping)
|
311 |
+
elif 'text/html' in content_type:
|
312 |
+
html_content = response.text
|
313 |
+
|
314 |
+
# Extract images from HTML
|
315 |
+
img_urls = re.findall(r'<img[^>]+src=["\']([^"\']+)["\']', html_content)
|
316 |
+
|
317 |
+
# Try to get the first valid image
|
318 |
+
for img_url in img_urls[:3]: # Try first 3 images
|
319 |
+
try:
|
320 |
+
if not img_url.startswith('http'):
|
321 |
+
img_url = requests.compat.urljoin(url, img_url)
|
322 |
+
|
323 |
+
img_response = requests.get(img_url, headers=headers, timeout=10)
|
324 |
+
img_response.raise_for_status()
|
325 |
+
|
326 |
+
image = Image.open(BytesIO(img_response.content))
|
327 |
+
|
328 |
+
# Extract text content from HTML
|
329 |
+
text_content = re.sub(r'<[^>]+>', ' ', html_content)
|
330 |
+
text_content = re.sub(r'\s+', ' ', text_content)[:500]
|
331 |
+
|
332 |
+
return {
|
333 |
+
"type": "webpage",
|
334 |
+
"content": image,
|
335 |
+
"text": text_content,
|
336 |
+
"url": url
|
337 |
+
}
|
338 |
+
|
339 |
+
except Exception as img_e:
|
340 |
+
logger.warning(f"Failed to fetch image {img_url}: {img_e}")
|
341 |
+
continue
|
342 |
+
|
343 |
+
# If no images found, return text content
|
344 |
+
text_content = re.sub(r'<[^>]+>', ' ', html_content)
|
345 |
+
text_content = re.sub(r'\s+', ' ', text_content)[:1000]
|
346 |
+
|
347 |
+
return {"type": "text", "content": text_content, "url": url}
|
348 |
+
|
349 |
+
else:
|
350 |
+
return {"type": "error", "content": f"Unsupported content type: {content_type}"}
|
351 |
+
|
352 |
+
except requests.RequestException as e:
|
353 |
+
logger.error(f"Request error for URL {url}: {e}")
|
354 |
+
return {"type": "error", "content": f"Failed to fetch URL: {str(e)}"}
|
355 |
+
except Exception as e:
|
356 |
+
logger.error(f"General error fetching {url}: {e}")
|
357 |
+
return {"type": "error", "content": f"Error processing content: {str(e)}"}
|
358 |
+
|
359 |
+
def ensemble_prediction(self, sentiment_result: Dict, multimodal_result: Dict, extracted_text: str = "") -> Dict:
|
360 |
+
"""Advanced ensemble prediction with risk stratification"""
|
361 |
+
|
362 |
+
# Convert sentiment to risk score
|
363 |
+
sentiment_risk = self.sentiment_to_risk_score(sentiment_result["label"], sentiment_result["score"])
|
364 |
+
|
365 |
+
# Get multimodal risk score
|
366 |
+
multimodal_risk = multimodal_result["hate_probability"]
|
367 |
+
|
368 |
+
# Context-aware weighting
|
369 |
+
text_weight = self.ensemble_weights['text_sentiment']
|
370 |
+
multimodal_weight = self.ensemble_weights['image_content'] + self.ensemble_weights['multimodal_context']
|
371 |
+
|
372 |
+
# Adjust weights based on text availability
|
373 |
+
if not extracted_text.strip():
|
374 |
+
text_weight *= 0.5
|
375 |
+
multimodal_weight = 1.0 - text_weight
|
376 |
+
|
377 |
+
# Calculate combined risk score
|
378 |
+
combined_risk = (text_weight * sentiment_risk + multimodal_weight * multimodal_risk)
|
379 |
+
|
380 |
+
# Risk stratification
|
381 |
+
if combined_risk >= self.risk_thresholds['high_risk']:
|
382 |
+
risk_level = "HIGH"
|
383 |
+
risk_description = "Potentially harmful content requiring immediate attention"
|
384 |
+
elif combined_risk >= self.risk_thresholds['medium_risk']:
|
385 |
+
risk_level = "MEDIUM"
|
386 |
+
risk_description = "Concerning content that may require review"
|
387 |
+
elif combined_risk >= self.risk_thresholds['low_risk']:
|
388 |
+
risk_level = "LOW"
|
389 |
+
risk_description = "Mildly concerning content, likely safe"
|
390 |
+
else:
|
391 |
+
risk_level = "SAFE"
|
392 |
+
risk_description = "Content appears safe and non-harmful"
|
393 |
+
|
394 |
+
# Confidence calculation
|
395 |
+
confidence = self.calculate_ensemble_confidence(sentiment_result, multimodal_result)
|
396 |
+
|
397 |
+
return {
|
398 |
+
"risk_level": risk_level,
|
399 |
+
"risk_score": combined_risk,
|
400 |
+
"risk_description": risk_description,
|
401 |
+
"confidence": confidence,
|
402 |
+
"sentiment_analysis": sentiment_result,
|
403 |
+
"multimodal_analysis": multimodal_result,
|
404 |
+
"explanation": self.generate_explanation(sentiment_result, multimodal_result, risk_level)
|
405 |
+
}
|
406 |
+
|
407 |
+
def sentiment_to_risk_score(self, sentiment_label: str, confidence: float) -> float:
|
408 |
+
"""Convert sentiment analysis to risk score"""
|
409 |
+
base_scores = {"NEGATIVE": 0.7, "NEUTRAL": 0.3, "POSITIVE": 0.1}
|
410 |
+
base_score = base_scores.get(sentiment_label, 0.3)
|
411 |
+
|
412 |
+
# Adjust based on confidence
|
413 |
+
return base_score * confidence + (1 - confidence) * 0.3
|
414 |
+
|
415 |
+
def calculate_ensemble_confidence(self, sentiment_result: Dict, multimodal_result: Dict) -> float:
|
416 |
+
"""Calculate overall ensemble confidence"""
|
417 |
+
sentiment_conf = sentiment_result["score"]
|
418 |
+
multimodal_conf = multimodal_result["confidence"]
|
419 |
+
|
420 |
+
# Weighted average of confidences
|
421 |
+
overall_conf = (sentiment_conf + multimodal_conf) / 2
|
422 |
+
|
423 |
+
# Boost confidence if both models agree
|
424 |
+
sentiment_negative = sentiment_result["label"] == "NEGATIVE"
|
425 |
+
multimodal_hateful = multimodal_result["is_hateful"]
|
426 |
+
|
427 |
+
if sentiment_negative == multimodal_hateful:
|
428 |
+
overall_conf = min(1.0, overall_conf * 1.2)
|
429 |
+
|
430 |
+
return overall_conf
|
431 |
+
|
432 |
+
def generate_explanation(self, sentiment_result: Dict, multimodal_result: Dict, risk_level: str) -> str:
|
433 |
+
"""Generate human-readable explanation of the decision"""
|
434 |
+
explanations = []
|
435 |
+
|
436 |
+
# Sentiment explanation
|
437 |
+
sentiment_label = sentiment_result["label"]
|
438 |
+
sentiment_conf = sentiment_result["score"]
|
439 |
+
explanations.append(f"Text sentiment: {sentiment_label} (confidence: {sentiment_conf:.1%})")
|
440 |
+
|
441 |
+
# Multimodal explanation
|
442 |
+
hate_prob = multimodal_result["hate_probability"]
|
443 |
+
explanations.append(f"Visual content analysis: {hate_prob:.1%} probability of harmful content")
|
444 |
+
|
445 |
+
# Risk level explanation
|
446 |
+
explanations.append(f"Overall risk assessment: {risk_level}")
|
447 |
+
|
448 |
+
return " | ".join(explanations)
|
449 |
+
|
450 |
+
# Initialize the analyzer
|
451 |
+
analyzer = EnhancedEnsembleMemeAnalyzer()
|
452 |
+
|
453 |
+
def analyze_content(input_type: str, text_input: str, image_input: Image.Image, url_input: str) -> Tuple[str, str, str]:
|
454 |
+
"""Main analysis function for Gradio interface"""
|
455 |
+
try:
|
456 |
+
extracted_text = ""
|
457 |
+
image_content = None
|
458 |
+
source_info = ""
|
459 |
+
|
460 |
+
# Handle different input types
|
461 |
+
if input_type == "Text Only" and text_input:
|
462 |
+
extracted_text = text_input
|
463 |
+
source_info = "Direct text input"
|
464 |
+
|
465 |
+
elif input_type == "Image Only" and image_input:
|
466 |
+
image_content = image_input
|
467 |
+
extracted_text = analyzer.extract_text_from_image(image_input)
|
468 |
+
source_info = "Direct image upload"
|
469 |
+
|
470 |
+
elif input_type == "URL" and url_input:
|
471 |
+
content = analyzer.fetch_social_media_content(url_input)
|
472 |
+
source_info = f"Content from: {url_input}"
|
473 |
+
|
474 |
+
if content["type"] == "image":
|
475 |
+
image_content = content["content"]
|
476 |
+
extracted_text = analyzer.extract_text_from_image(content["content"])
|
477 |
+
elif content["type"] == "webpage":
|
478 |
+
image_content = content["content"]
|
479 |
+
extracted_text = content.get("text", "") + " " + analyzer.extract_text_from_image(content["content"])
|
480 |
+
elif content["type"] == "text":
|
481 |
+
extracted_text = content["content"]
|
482 |
+
else:
|
483 |
+
return f"β Error: {content['content']}", "", ""
|
484 |
+
|
485 |
+
elif input_type == "Text + Image" and text_input and image_input:
|
486 |
+
extracted_text = text_input + " " + analyzer.extract_text_from_image(image_input)
|
487 |
+
image_content = image_input
|
488 |
+
source_info = "Combined text and image input"
|
489 |
+
|
490 |
+
else:
|
491 |
+
return "β οΈ Please provide appropriate input based on the selected type.", "", ""
|
492 |
+
|
493 |
+
# Perform analysis
|
494 |
+
sentiment_result = analyzer.analyze_sentiment(extracted_text)
|
495 |
+
|
496 |
+
if image_content:
|
497 |
+
multimodal_result = analyzer.classify_multimodal_content(image_content, extracted_text)
|
498 |
+
else:
|
499 |
+
# Default multimodal analysis for text-only content
|
500 |
+
multimodal_result = {
|
501 |
+
"is_hateful": False,
|
502 |
+
"hate_probability": 0.2,
|
503 |
+
"safe_probability": 0.8,
|
504 |
+
"confidence": 0.5,
|
505 |
+
"detailed_scores": []
|
506 |
+
}
|
507 |
+
|
508 |
+
# Get ensemble prediction
|
509 |
+
final_result = analyzer.ensemble_prediction(sentiment_result, multimodal_result, extracted_text)
|
510 |
+
|
511 |
+
# Format comprehensive results
|
512 |
+
risk_emoji = {"HIGH": "π¨", "MEDIUM": "β οΈ", "LOW": "π‘", "SAFE": "β
"}
|
513 |
+
|
514 |
+
result_text = f"""
|
515 |
+
# π€ Enhanced Ensemble Analysis Results
|
516 |
+
|
517 |
+
## {risk_emoji[final_result['risk_level']]} Overall Assessment
|
518 |
+
**Risk Level**: {final_result['risk_level']}
|
519 |
+
**Risk Score**: {final_result['risk_score']:.1%}
|
520 |
+
**Confidence**: {final_result['confidence']:.1%}
|
521 |
+
**Description**: {final_result['risk_description']}
|
522 |
+
|
523 |
+
---
|
524 |
+
|
525 |
+
## π Detailed Analysis
|
526 |
+
|
527 |
+
### π Text Analysis
|
528 |
+
**Source**: {source_info}
|
529 |
+
**Extracted Text**: {extracted_text[:300]}{'...' if len(extracted_text) > 300 else ''}
|
530 |
+
**Sentiment**: {sentiment_result['label']} ({sentiment_result['score']:.1%} confidence)
|
531 |
+
|
532 |
+
### πΌοΈ Visual Content Analysis
|
533 |
+
**Contains Harmful Content**: {'Yes' if multimodal_result['is_hateful'] else 'No'}
|
534 |
+
**Harm Probability**: {multimodal_result['hate_probability']:.1%}
|
535 |
+
**Safe Probability**: {multimodal_result['safe_probability']:.1%}
|
536 |
+
**Visual Analysis Confidence**: {multimodal_result['confidence']:.1%}
|
537 |
+
|
538 |
+
### π§ Ensemble Decision Process
|
539 |
+
{final_result['explanation']}
|
540 |
+
|
541 |
+
---
|
542 |
+
|
543 |
+
## π‘ Recommendations
|
544 |
+
{analyzer.get_recommendations(final_result['risk_level'])}
|
545 |
+
"""
|
546 |
+
|
547 |
+
# Prepare detailed output for inspection
|
548 |
+
detailed_output = json.dumps({
|
549 |
+
"risk_assessment": {
|
550 |
+
"level": final_result['risk_level'],
|
551 |
+
"score": final_result['risk_score'],
|
552 |
+
"confidence": final_result['confidence']
|
553 |
+
},
|
554 |
+
"text_analysis": sentiment_result,
|
555 |
+
"visual_analysis": multimodal_result,
|
556 |
+
"extracted_text": extracted_text
|
557 |
+
}, indent=2)
|
558 |
+
|
559 |
+
return result_text, extracted_text, detailed_output
|
560 |
+
|
561 |
+
except Exception as e:
|
562 |
+
logger.error(f"Analysis error: {e}")
|
563 |
+
return f"β Error during analysis: {str(e)}", "", ""
|
564 |
+
|
565 |
+
# Add recommendation method to analyzer class
|
566 |
+
def get_recommendations(self, risk_level: str) -> str:
|
567 |
+
"""Get recommendations based on risk level"""
|
568 |
+
recommendations = {
|
569 |
+
"HIGH": "π¨ **Immediate Action Required**: This content should be reviewed by moderators and potentially removed. Consider issuing warnings or taking enforcement action.",
|
570 |
+
"MEDIUM": "β οΈ **Review Recommended**: Content may violate community guidelines. Manual review suggested before taking action.",
|
571 |
+
"LOW": "π‘ **Monitor**: Content shows some concerning signals but may be acceptable. Consider additional context before action.",
|
572 |
+
"SAFE": "β
**No Action Needed**: Content appears safe and compliant with community standards."
|
573 |
+
}
|
574 |
+
return recommendations.get(risk_level, "No specific recommendations available.")
|
575 |
+
|
576 |
+
# Add the method to the class
|
577 |
+
EnhancedEnsembleMemeAnalyzer.get_recommendations = get_recommendations
|
578 |
+
|
579 |
+
# Create enhanced Gradio interface
|
580 |
+
with gr.Blocks(title="Enhanced Ensemble Meme & Text Analyzer", theme=gr.themes.Soft()) as demo:
|
581 |
+
gr.Markdown("""
|
582 |
+
# π€ Enhanced Ensemble Meme & Text Analyzer
|
583 |
+
|
584 |
+
**Advanced AI system combining:**
|
585 |
+
- π― Fine-tuned BERT (93% accuracy) for sentiment analysis
|
586 |
+
- ποΈ SigLIP-Large for visual content understanding
|
587 |
+
- π Advanced OCR for text extraction
|
588 |
+
- π§ Intelligent ensemble decision making
|
589 |
+
|
590 |
+
**Analyzes content risk across multiple dimensions with explainable AI**
|
591 |
+
""")
|
592 |
+
|
593 |
+
with gr.Row():
|
594 |
+
input_type = gr.Dropdown(
|
595 |
+
choices=["Text Only", "Image Only", "URL", "Text + Image"],
|
596 |
+
value="Text Only",
|
597 |
+
label="π₯ Input Type",
|
598 |
+
info="Select the type of content you want to analyze"
|
599 |
+
)
|
600 |
+
|
601 |
+
with gr.Row():
|
602 |
+
with gr.Column(scale=2):
|
603 |
+
text_input = gr.Textbox(
|
604 |
+
label="π Text Input",
|
605 |
+
placeholder="Enter text content to analyze (tweets, posts, comments)...",
|
606 |
+
lines=4
|
607 |
+
)
|
608 |
+
image_input = gr.Image(
|
609 |
+
label="πΌοΈ Image Input",
|
610 |
+
type="pil",
|
611 |
+
info="Upload memes, screenshots, or social media images"
|
612 |
+
)
|
613 |
+
url_input = gr.Textbox(
|
614 |
+
label="π URL Input",
|
615 |
+
placeholder="Enter social media URL (Twitter, Reddit, etc.)...",
|
616 |
+
info="Paste links to posts, images, or web content"
|
617 |
+
)
|
618 |
+
|
619 |
+
with gr.Column(scale=1):
|
620 |
+
analyze_btn = gr.Button("π Analyze Content", variant="primary", size="lg")
|
621 |
+
|
622 |
+
gr.Markdown("""
|
623 |
+
### π― Model Information
|
624 |
+
- **BERT**: Fine-tuned sentiment analysis (93% accuracy)
|
625 |
+
- **SigLIP**: Large-scale vision-language model
|
626 |
+
- **OCR**: Multi-engine text extraction
|
627 |
+
- **Ensemble**: Weighted decision fusion
|
628 |
+
""")
|
629 |
+
|
630 |
+
with gr.Row():
|
631 |
+
output_analysis = gr.Markdown(label="π Analysis Results")
|
632 |
+
|
633 |
+
with gr.Row():
|
634 |
+
with gr.Column():
|
635 |
+
output_text = gr.Textbox(label="π Extracted Text", lines=4)
|
636 |
+
with gr.Column():
|
637 |
+
output_detailed = gr.Code(label="π§ Detailed Results (JSON)", language="json")
|
638 |
+
|
639 |
+
# Enhanced examples
|
640 |
+
gr.Examples(
|
641 |
+
examples=[
|
642 |
+
["Text Only", "This meme is so offensive and targets innocent people. Absolutely disgusting!", None, ""],
|
643 |
+
["Text Only", "Haha this meme made my day! So funny and clever π", None, ""],
|
644 |
+
["URL", "", None, "https://i.imgur.com/example.jpg"],
|
645 |
+
["Text + Image", "Check out this hilarious meme I found!", None, ""]
|
646 |
+
],
|
647 |
+
inputs=[input_type, text_input, image_input, url_input],
|
648 |
+
label="π‘ Try these examples"
|
649 |
+
)
|
650 |
+
|
651 |
+
analyze_btn.click(
|
652 |
+
fn=analyze_content,
|
653 |
+
inputs=[input_type, text_input, image_input, url_input],
|
654 |
+
outputs=[output_analysis, output_text, output_detailed]
|
655 |
+
)
|
656 |
+
|
657 |
+
if __name__ == "__main__":
|
658 |
+
demo.launch(
|
659 |
+
share=True,
|
660 |
+
server_name="0.0.0.0",
|
661 |
+
server_port=7860,
|
662 |
+
show_error=True
|
663 |
+
)
|