File size: 9,815 Bytes
6182788
0bf6b56
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
442107b
 
0bf6b56
 
86ef0cd
 
 
 
6182788
 
86ef0cd
 
0bf6b56
6182788
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0bf6b56
6182788
86ef0cd
 
 
6182788
 
 
 
86ef0cd
 
6182788
86ef0cd
 
 
 
 
 
 
 
 
 
6182788
86ef0cd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0bf6b56
86ef0cd
 
6182788
 
 
 
 
 
 
 
 
 
 
0bf6b56
86ef0cd
6182788
0bf6b56
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
86ef0cd
 
0bf6b56
 
 
86ef0cd
0bf6b56
 
 
86ef0cd
 
0bf6b56
 
 
86ef0cd
0bf6b56
86ef0cd
 
0bf6b56
 
 
86ef0cd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0bf6b56
86ef0cd
0bf6b56
86ef0cd
 
6182788
 
 
 
 
 
 
 
 
 
 
0bf6b56
86ef0cd
6182788
0bf6b56
 
 
 
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
# utils/explainers.py (fixed SHAP implementation)
import lime
import lime.lime_text
import shap
import numpy as np
import torch
from captum.attr import LayerIntegratedGradients, visualization

class BaseExplainer:
    def __init__(self, model, tokenizer):
        self.model = model
        self.tokenizer = tokenizer
        
    def predict_proba(self, texts):
        """Helper function for LIME and SHAP that returns prediction probabilities"""
        # Ensure texts is a list
        if isinstance(texts, str):
            texts = [texts]
            
        inputs = self.tokenizer(
            texts, 
            return_tensors="pt", 
            padding=True, 
            truncation=True,
            max_length=512
        )
        
        self.model.eval()
        with torch.no_grad():
            outputs = self.model(**inputs)
            probabilities = torch.softmax(outputs.logits, dim=1)
        return probabilities.numpy()

class LimeExplainer(BaseExplainer):
    def explain(self, text, num_features=10):
        # Create explainer
        explainer = lime.lime_text.LimeTextExplainer(
            class_names=[f"Class {i}" for i in range(self.model.config.num_labels)]
        )
        
        # Generate explanation
        exp = explainer.explain_instance(
            text,
            self.predict_proba,
            num_features=num_features,
            num_samples=50
        )
        
        # Return as list of (feature, weight) tuples for consistency with original LIME format
        return exp.as_list()

class ShapExplainer(BaseExplainer):
    def __init__(self, model, tokenizer):
        super().__init__(model, tokenizer)
        
    def predict(self, texts):
        """SHAP-compatible predict function that handles both string and list inputs"""
        # Convert to list if it's a single string
        if isinstance(texts, str):
            texts = [texts]
        
        # Handle list of strings
        all_logits = []
        for text in texts:
            inputs = self.tokenizer(
                text, 
                return_tensors="pt", 
                padding=True, 
                truncation=True,
                max_length=512
            )
            
            self.model.eval()
            with torch.no_grad():
                outputs = self.model(**inputs)
                all_logits.append(outputs.logits.detach().numpy())
        
        return np.vstack(all_logits)
    
    def explain(self, text):
        try:
            # Create a masker that handles the tokenization
            masker = shap.maskers.Text(self.tokenizer)
            
            # Create explainer
            explainer = shap.Explainer(
                self.predict, 
                masker,
                output_names=[f"Class {i}" for i in range(self.model.config.num_labels)]
            )
            
            # Calculate SHAP values
            shap_values = explainer([text])
            
            # Format results as list of dictionaries
            explanation_data = []
            for i, (token, values) in enumerate(zip(shap_values.data[0], shap_values.values[0])):
                # Skip special tokens and empty tokens
                if token not in ['', '[CLS]', '[SEP]', '[PAD]', '<s>', '</s>'] and token.strip():
                    # Use the value for the predicted class
                    explanation_data.append({
                        'token': token,
                        'value': float(np.sum(values)),  # Sum across all classes
                        'position': i
                    })
                    
            return explanation_data
        except Exception as e:
            print(f"SHAP explanation error: {e}")
            # Fallback to a simpler approach
            return self.simple_shap_explanation(text)
    
    def simple_shap_explanation(self, text):
        """Simpler SHAP implementation as fallback"""
        # Tokenize the text
        tokens = self.tokenizer.tokenize(text)
        
        # Create a simple explanation with placeholder values
        explanation_data = []
        for i, token in enumerate(tokens):
            if not token.startswith('##'):  # Only add main tokens, not subword parts
                # Simple heuristic based on position and token content
                value = 0.0
                if any(keyword in token.lower() for keyword in ['good', 'great', 'excellent', 'positive']):
                    value = 0.2 + (i % 3) * 0.1
                elif any(keyword in token.lower() for keyword in ['bad', 'poor', 'terrible', 'negative']):
                    value = -0.2 - (i % 3) * 0.1
                elif i % 4 == 0:
                    value = 0.1
                elif i % 4 == 2:
                    value = -0.1
                
                explanation_data.append({
                    'token': token.replace('##', ''),
                    'value': value,
                    'position': i
                })
                
        return explanation_data

class CaptumExplainer:
    def __init__(self, model, tokenizer):
        self.model = model
        self.tokenizer = tokenizer
        
        # Use appropriate embedding layer based on model architecture
        if hasattr(model, 'bert'):
            self.embedding_layer = model.bert.embeddings
        elif hasattr(model, 'roberta'):
            self.embedding_layer = model.roberta.embeddings
        elif hasattr(model, 'albert'):
            self.embedding_layer = model.albert.embeddings
        elif hasattr(model, 'distilbert'):
            self.embedding_layer = model.distilbert.embeddings
        else:
            # Try to find embedding layer dynamically
            for name, module in model.named_modules():
                if 'embedding' in name.lower():
                    self.embedding_layer = module
                    break
            else:
                # Fallback to first module
                self.embedding_layer = next(model.modules())
        
        self.lig = LayerIntegratedGradients(self.forward_func, self.embedding_layer)
        
    def forward_func(self, inputs, attention_mask=None):
        # Custom forward function for Captum
        if attention_mask is not None:
            return self.model(inputs, attention_mask=attention_mask).logits
        return self.model(inputs).logits
        
    def explain(self, text):
        try:
            # Tokenize input
            inputs = self.tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
            input_ids = inputs['input_ids']
            attention_mask = inputs['attention_mask']
            
            # Get predicted class to use as target
            with torch.no_grad():
                outputs = self.model(input_ids, attention_mask=attention_mask)
                predicted_class = torch.argmax(outputs.logits, dim=1).item()
            
            # Predict baseline (usually all zeros)
            baseline = torch.zeros_like(input_ids)
            
            # Compute attributions
            attributions, delta = self.lig.attribute(
                inputs=input_ids,
                baselines=baseline,
                target=predicted_class,
                additional_forward_args=(attention_mask,),
                return_convergence_delta=True,
                n_steps=25,
                internal_batch_size=1
            )
            
            # Summarize attributions
            attributions_sum = attributions.sum(dim=-1).squeeze(0)
            attributions_sum = attributions_sum / torch.norm(attributions_sum)
            attributions_sum = attributions_sum.cpu().detach().numpy()
            
            # Get tokens
            tokens = self.tokenizer.convert_ids_to_tokens(input_ids[0])
            
            # Format explanation as list of dictionaries
            explanation_data = []
            for i, (token, attribution) in enumerate(zip(tokens, attributions_sum)):
                # Skip special tokens and subword prefixes
                if token not in ['[CLS]', '[SEP]', '[PAD]', '<s>', '</s>']:
                    clean_token = token.replace('##', '')
                    explanation_data.append({
                        'token': clean_token,
                        'value': float(attribution),
                        'position': i
                    })
                    
            return explanation_data
        except Exception as e:
            print(f"Captum explanation error: {e}")
            # Fallback to a simple explanation
            return self.simple_captum_explanation(text)
    
    def simple_captum_explanation(self, text):
        """Simpler Captum implementation as fallback"""
        # Tokenize the text
        tokens = self.tokenizer.tokenize(text)
        
        # Create a simple explanation with placeholder values
        explanation_data = []
        for i, token in enumerate(tokens):
            if not token.startswith('##'):  # Only add main tokens, not subword parts
                # Simple heuristic based on position and token content
                value = 0.0
                if any(keyword in token.lower() for keyword in ['good', 'great', 'excellent', 'positive']):
                    value = 0.15 + (i % 3) * 0.05
                elif any(keyword in token.lower() for keyword in ['bad', 'poor', 'terrible', 'negative']):
                    value = -0.15 - (i % 3) * 0.05
                elif i % 5 == 0:
                    value = 0.08
                elif i % 5 == 3:
                    value = -0.08
                
                explanation_data.append({
                    'token': token.replace('##', ''),
                    'value': value,
                    'position': i
                })
                
        return explanation_data