File size: 16,055 Bytes
217a100 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 |
import torch
import torch.nn.functional as F
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
from transformers import LlamaTokenizer, LlamaForCausalLM
import pandas as pd
from sklearn.metrics.pairwise import cosine_similarity
import lime
from lime.lime_text import LimeTextExplainer
import shap
import re
import warnings
warnings.filterwarnings('ignore')
class LLMExplainabilityAnalyzer:
def __init__(self, model_path, tokenizer_path=None):
"""Initialize with model and tokenizer paths"""
self.model_path = model_path
self.tokenizer_path = tokenizer_path or model_path
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Load model and tokenizer
self.load_model()
# Initialize explanation tools
self.lime_explainer = LimeTextExplainer(class_names=['Generated Text'])
def load_model(self):
"""Load the fine-tuned model and tokenizer"""
try:
print(f"Loading model from: {self.model_path}")
self.tokenizer = AutoTokenizer.from_pretrained(self.tokenizer_path)
self.model = AutoModelForCausalLM.from_pretrained(
self.model_path,
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
device_map="auto" if torch.cuda.is_available() else None
)
# Set padding token if not exists
if self.tokenizer.pad_token is None:
self.tokenizer.pad_token = self.tokenizer.eos_token
print("Model loaded successfully!")
except Exception as e:
print(f"Error loading model: {e}")
# Fallback to base model
print("Loading base TinyLlama model...")
self.tokenizer = AutoTokenizer.from_pretrained("TinyLlama/TinyLlama-1.1B-Chat-v1.0")
self.model = AutoModelForCausalLM.from_pretrained("TinyLlama/TinyLlama-1.1B-Chat-v1.0")
if self.tokenizer.pad_token is None:
self.tokenizer.pad_token = self.tokenizer.eos_token
def extract_attention_weights(self, text, max_length=512):
"""Extract attention weights for visualization"""
inputs = self.tokenizer(
text,
return_tensors="pt",
max_length=max_length,
truncation=True,
padding=True
).to(self.device)
with torch.no_grad():
outputs = self.model(**inputs, output_attentions=True)
attentions = outputs.attentions
# Get tokens
tokens = self.tokenizer.convert_ids_to_tokens(inputs['input_ids'][0])
return attentions, tokens
def visualize_attention_heads(self, text, layer_idx=0, head_idx=0, max_length=512):
"""Visualize attention patterns for specific layer and head"""
attentions, tokens = self.extract_attention_weights(text, max_length)
# Get attention weights for specific layer and head
attention_weights = attentions[layer_idx][0, head_idx].cpu().numpy()
# Create heatmap
plt.figure(figsize=(12, 8))
sns.heatmap(
attention_weights,
xticklabels=tokens,
yticklabels=tokens,
cmap='Blues',
cbar=True
)
plt.title(f'Attention Weights - Layer {layer_idx}, Head {head_idx}')
plt.xlabel('Key Tokens')
plt.ylabel('Query Tokens')
plt.xticks(rotation=45)
plt.yticks(rotation=0)
plt.tight_layout()
plt.show()
return attention_weights, tokens
def attention_rollout(self, text, max_length=512):
"""Compute attention rollout for global attention patterns"""
attentions, tokens = self.extract_attention_weights(text, max_length)
# Convert to numpy
attention_matrices = [att[0].mean(dim=0).cpu().numpy() for att in attentions]
# Compute rollout
rollout = attention_matrices[0]
for attention_matrix in attention_matrices[1:]:
rollout = np.matmul(rollout, attention_matrix)
# Visualize rollout
plt.figure(figsize=(12, 8))
sns.heatmap(
rollout,
xticklabels=tokens,
yticklabels=tokens,
cmap='Reds',
cbar=True
)
plt.title('Attention Rollout - Global Attention Flow')
plt.xlabel('Key Tokens')
plt.ylabel('Query Tokens')
plt.xticks(rotation=45)
plt.yticks(rotation=0)
plt.tight_layout()
plt.show()
return rollout, tokens
def gradient_saliency(self, text, target_token_idx=None, max_length=512):
"""Compute gradient-based saliency maps"""
inputs = self.tokenizer(
text,
return_tensors="pt",
max_length=max_length,
truncation=True,
padding=True
).to(self.device)
# Enable gradients for embeddings
embeddings = self.model.get_input_embeddings()
inputs_embeds = embeddings(inputs['input_ids'])
inputs_embeds.requires_grad_(True)
# Forward pass
outputs = self.model(inputs_embeds=inputs_embeds, attention_mask=inputs['attention_mask'])
# Get target logits (last token if not specified)
if target_token_idx is None:
target_token_idx = -1
target_logits = outputs.logits[0, target_token_idx]
target_prob = F.softmax(target_logits, dim=-1)
# Compute gradients
target_prob.max().backward()
# Get saliency scores
saliency_scores = inputs_embeds.grad.norm(dim=-1).squeeze().cpu().numpy()
tokens = self.tokenizer.convert_ids_to_tokens(inputs['input_ids'][0])
# Visualize saliency
plt.figure(figsize=(15, 6))
colors = plt.cm.Reds(saliency_scores / saliency_scores.max())
for i, (token, score) in enumerate(zip(tokens, saliency_scores)):
plt.bar(i, score, color=colors[i])
plt.text(i, score + 0.001, token, rotation=45, ha='left', va='bottom')
plt.title('Gradient Saliency Scores')
plt.xlabel('Token Position')
plt.ylabel('Saliency Score')
plt.tight_layout()
plt.show()
return saliency_scores, tokens
def lime_explanation(self, text, num_samples=1000):
"""Generate LIME explanations"""
def predict_fn(texts):
"""Prediction function for LIME"""
predictions = []
for text in texts:
try:
inputs = self.tokenizer(
text,
return_tensors="pt",
max_length=512,
truncation=True,
padding=True
).to(self.device)
with torch.no_grad():
outputs = self.model(**inputs)
logits = outputs.logits[0, -1]
probs = F.softmax(logits, dim=-1)
# Return probability distribution
predictions.append(probs.cpu().numpy())
except:
# Return uniform distribution if error
predictions.append(np.ones(self.tokenizer.vocab_size) / self.tokenizer.vocab_size)
return np.array(predictions)
# Generate explanation
explanation = self.lime_explainer.explain_instance(
text,
predict_fn,
num_features=20,
num_samples=num_samples
)
# Visualize explanation
explanation.show_in_notebook(text=True)
return explanation
def activation_analysis(self, text, layer_indices=None, max_length=512):
"""Analyze hidden layer activations"""
inputs = self.tokenizer(
text,
return_tensors="pt",
max_length=max_length,
truncation=True,
padding=True
).to(self.device)
# Hook to capture activations
activations = {}
def hook_fn(name):
def hook(module, input, output):
activations[name] = output.detach()
return hook
# Register hooks
if layer_indices is None:
layer_indices = [0, len(self.model.model.layers)//2, len(self.model.model.layers)-1]
hooks = []
for idx in layer_indices:
if idx < len(self.model.model.layers):
hook = self.model.model.layers[idx].register_forward_hook(hook_fn(f'layer_{idx}'))
hooks.append(hook)
# Forward pass
with torch.no_grad():
outputs = self.model(**inputs)
# Remove hooks
for hook in hooks:
hook.remove()
# Analyze activations
tokens = self.tokenizer.convert_ids_to_tokens(inputs['input_ids'][0])
for layer_name, activation in activations.items():
# Get activation statistics
activation_np = activation[0].cpu().numpy()
# Plot activation distribution
plt.figure(figsize=(12, 6))
# Heatmap of activations
plt.subplot(1, 2, 1)
sns.heatmap(activation_np.T, cmap='viridis', cbar=True)
plt.title(f'{layer_name} Activations')
plt.xlabel('Token Position')
plt.ylabel('Hidden Dimension')
# Activation magnitude per token
plt.subplot(1, 2, 2)
activation_magnitudes = np.linalg.norm(activation_np, axis=1)
plt.bar(range(len(tokens)), activation_magnitudes)
plt.title(f'{layer_name} Activation Magnitudes')
plt.xlabel('Token Position')
plt.ylabel('Magnitude')
plt.xticks(range(len(tokens)), tokens, rotation=45)
plt.tight_layout()
plt.show()
def token_importance_analysis(self, text, method='attention', max_length=512):
"""Analyze token importance using different methods"""
results = {}
if method == 'attention':
# Attention-based importance
attentions, tokens = self.extract_attention_weights(text, max_length)
# Average attention across layers and heads
avg_attention = torch.stack([att.mean(dim=1) for att in attentions]).mean(dim=0)
importance_scores = avg_attention[0].sum(dim=0).cpu().numpy()
elif method == 'gradient':
# Gradient-based importance
importance_scores, tokens = self.gradient_saliency(text, max_length=max_length)
# Create importance dataframe
importance_df = pd.DataFrame({
'token': tokens,
'importance': importance_scores
})
# Sort by importance
importance_df = importance_df.sort_values('importance', ascending=False)
# Visualize top important tokens
plt.figure(figsize=(12, 6))
top_tokens = importance_df.head(20)
plt.barh(range(len(top_tokens)), top_tokens['importance'])
plt.yticks(range(len(top_tokens)), top_tokens['token'])
plt.title(f'Top 20 Important Tokens ({method.title()} Method)')
plt.xlabel('Importance Score')
plt.tight_layout()
plt.show()
return importance_df
def semantic_similarity_analysis(self, texts, max_length=512):
"""Analyze semantic similarity between different texts"""
embeddings = []
for text in texts:
inputs = self.tokenizer(
text,
return_tensors="pt",
max_length=max_length,
truncation=True,
padding=True
).to(self.device)
with torch.no_grad():
outputs = self.model(**inputs, output_hidden_states=True)
# Use last layer, last token embedding
embedding = outputs.hidden_states[-1][0, -1].cpu().numpy()
embeddings.append(embedding)
# Compute similarity matrix
similarity_matrix = cosine_similarity(embeddings)
# Visualize similarity matrix
plt.figure(figsize=(10, 8))
sns.heatmap(
similarity_matrix,
annot=True,
cmap='viridis',
xticklabels=[f'Text {i+1}' for i in range(len(texts))],
yticklabels=[f'Text {i+1}' for i in range(len(texts))]
)
plt.title('Semantic Similarity Matrix')
plt.tight_layout()
plt.show()
return similarity_matrix
def generate_explanation_report(self, text, output_file='xai_report.html'):
"""Generate comprehensive explanation report"""
print("Generating comprehensive XAI report...")
# Run all analyses
print("1. Extracting attention patterns...")
attention_weights, tokens = self.visualize_attention_heads(text)
print("2. Computing attention rollout...")
rollout, _ = self.attention_rollout(text)
print("3. Calculating gradient saliency...")
saliency_scores, _ = self.gradient_saliency(text)
print("4. Analyzing activations...")
self.activation_analysis(text)
print("5. Computing token importance...")
importance_df = self.token_importance_analysis(text)
# Create summary
print("\n=== XAI ANALYSIS SUMMARY ===")
print(f"Input text: {text[:100]}...")
print(f"Number of tokens: {len(tokens)}")
print(f"Most important tokens: {importance_df.head(5)['token'].tolist()}")
print(f"Average attention entropy: {np.mean(-np.sum(attention_weights * np.log(attention_weights + 1e-10), axis=1)):.4f}")
return {
'attention_weights': attention_weights,
'rollout': rollout,
'saliency_scores': saliency_scores,
'importance_df': importance_df,
'tokens': tokens
}
def main():
"""Main function to run XAI analysis"""
# Initialize analyzer (adjust model path as needed)
try:
analyzer = LLMExplainabilityAnalyzer("./fine_tuned_model")
except:
print("Fine-tuned model not found. Using base model for demonstration.")
analyzer = LLMExplainabilityAnalyzer("TinyLlama/TinyLlama-1.1B-Chat-v1.0")
# Sample skin disease text for analysis
sample_text = """
Patient presents with erythematous scaly patches on the elbows and knees,
consistent with psoriasis. The condition appears to be chronic with periods
of exacerbation. Treatment options include topical corticosteroids and
phototherapy for mild to moderate cases.
"""
print("Starting XAI Analysis...")
print("=" * 50)
# Generate comprehensive report
results = analyzer.generate_explanation_report(sample_text)
# Additional analyses
print("\n6. Semantic similarity analysis...")
test_texts = [
"Psoriasis treatment with topical corticosteroids",
"Eczema management using moisturizers",
"Melanoma diagnosis and surgical intervention"
]
similarity_matrix = analyzer.semantic_similarity_analysis(test_texts)
print("\n" + "=" * 50)
print("XAI ANALYSIS COMPLETE")
print("=" * 50)
return results
if __name__ == "__main__":
main()
|