File size: 6,426 Bytes
993b547
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from transformers import RobertaTokenizer, RobertaForMaskedLM
import torch
import torch.nn.functional as F
from models import TransformerVisualizer
from transformers import (
    RobertaForMaskedLM, RobertaForSequenceClassification, RobertaForQuestionAnswering,
)

class RoBERTaVisualizer(TransformerVisualizer):
    def __init__(self, task):
        super().__init__()
        self.task = task
        self.tokenizer = RobertaTokenizer.from_pretrained('roberta-base')
        if self.task == 'mlm':
            self.model = RobertaForMaskedLM.from_pretrained("roberta-base")
        elif self.task == 'sst':
            self.model = RobertaForSequenceClassification.from_pretrained('textattack/roberta-base-SST-2')
        elif self.task == 'mnli':
            self.model = RobertaForSequenceClassification.from_pretrained("roberta-large-mnli")


        self.model.to(self.device)
        self.model.eval()
        self.num_attention_layers = self.model.config.num_hidden_layers


    def tokenize(self, text, hypothesis = ''):

         

        if len(hypothesis) == 0:
            encoded = self.tokenizer(text, return_tensors='pt', return_attention_mask=True,padding=False, truncation=True)
        else:
            encoded = self.tokenizer(text, hypothesis, return_tensors='pt', return_attention_mask=True,padding=False, truncation=True)

        input_ids = encoded['input_ids'].to(self.device)
        attention_mask = encoded['attention_mask'].to(self.device)
        tokens = self.tokenizer.convert_ids_to_tokens(input_ids[0])
        print('First time tokenizing:', tokens, len(tokens))

        response = {
            'input_ids': input_ids,
            'attention_mask': attention_mask,
            'tokens': tokens
        }
        print(response)
        return response

    def predict(self, task, text, hypothesis='', maskID = None):
        
        

        if task == 'mlm':
            inputs = self.tokenizer(text, return_tensors='pt', padding=False, truncation=True)
            if maskID is not None and 0 <= maskID < inputs['input_ids'].size(1):
                inputs['input_ids'][0][maskID] = self.tokenizer.mask_token_id
                mask_index = maskID
            else:
                raise ValueError(f"Invalid maskID {maskID} for input of length {inputs['input_ids'].size(1)}")
            inputs = {k: v.to(self.device) for k, v in inputs.items()}

            with torch.no_grad():
                outputs = self.model(**inputs)
                logits = outputs.logits

            mask_logits = logits[0, mask_index]
            top_probs, top_indices = torch.topk(F.softmax(mask_logits, dim=-1), 10)
            decoded = self.tokenizer.convert_ids_to_tokens(top_indices.tolist())
            return decoded, top_probs

        elif task == 'sst':
            inputs = self.tokenizer(text, return_tensors='pt', padding=False, truncation=True).to(self.device)

            with torch.no_grad():
                outputs = self.model(**inputs)
                logits = outputs.logits
                probs = F.softmax(logits, dim=1).squeeze()

            labels = ["negative", "positive"]
            return labels, probs
        
        elif task == 'mnli': 
            inputs = self.tokenizer(text, hypothesis, return_tensors='pt', padding=True, truncation=True).to(self.device)

            with torch.no_grad():
                outputs = self.model(**inputs)
                logits = outputs.logits
                probs = F.softmax(logits, dim=1).squeeze()

            labels = ["entailment", "neutral", "contradiction"]
            return labels, probs

        else:
            raise NotImplementedError(f"Task '{task}' not supported for RoBERTa")


    def get_all_grad_attn_matrix(self, task, sentence, hypothesis='', maskID = None):
        print(task, sentence,  hypothesis)
        print('Tokenize')
        if task == 'mnli':
            inputs = self.tokenizer(sentence, hypothesis, return_tensors='pt', padding=False, truncation=True)
        elif task == 'mlm':
            inputs = self.tokenizer(sentence,  return_tensors='pt', padding=False, truncation=True)
            if maskID is not None and 0 <= maskID < inputs['input_ids'].size(1):
                inputs['input_ids'][0][maskID] = self.tokenizer.mask_token_id
        else:
            inputs = self.tokenizer(sentence,  return_tensors='pt', padding=False, truncation=True)
        tokens = self.tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
        print(tokens)
        inputs = {k: v.to(self.device) for k, v in inputs.items()}

        print('Input embeddings with grad')
        embedding_layer = self.model.roberta.embeddings.word_embeddings
        inputs_embeds = embedding_layer(inputs["input_ids"])
        inputs_embeds.requires_grad_()

        print('Forward pass')
        outputs = self.model.roberta(
            inputs_embeds=inputs_embeds,
            attention_mask=inputs["attention_mask"],
            output_attentions=True
        )
        attentions = outputs.attentions  # list of [1, heads, seq, seq]

        print('Average attentions per layer')
        mean_attns = [a.squeeze(0).mean(dim=0).detach().cpu() for a in attentions]

        attn_matrices_all = []
        grad_matrices_all = []
        for target_layer in range(len(attentions)):
            grad_matrix, attn_matrix = self.get_grad_attn_matrix(inputs_embeds, attentions, mean_attns, target_layer)
            grad_matrices_all.append(grad_matrix.tolist())
            attn_matrices_all.append(attn_matrix.tolist())
        return grad_matrices_all, attn_matrices_all
    
    def get_grad_attn_matrix(self,inputs_embeds, attentions, mean_attns, target_layer):
        
        attn_matrix = mean_attns[target_layer]
        seq_len = attn_matrix.shape[0]
        attn_layer = attentions[target_layer].squeeze(0).mean(dim=0)  # [seq, seq]

        print('Computing grad norms')
        grad_norms_list = []
        for k in range(seq_len):
            scalar = attn_layer[:, k].sum()
            grad = torch.autograd.grad(scalar, inputs_embeds, retain_graph=True)[0].squeeze(0)
            grad_norms = grad.norm(dim=1)
            grad_norms_list.append(grad_norms.unsqueeze(1))

        grad_matrix = torch.cat(grad_norms_list, dim=1)
        grad_matrix = grad_matrix[:seq_len, :seq_len]
        attn_matrix = attn_matrix[:seq_len, :seq_len]

        
        return grad_matrix, attn_matrix