CodeAttack/important_words.py at main · reddy-lab-code-research/CodeAttack · GitHub

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
import os, sys
parent_dir = os.path.abspath(os.path.dirname(__file__))
sys.path.append(parent_dir)

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader, SequentialSampler, TensorDataset

from graphcodebert_input import convert_examples_to_features

def _get_masked(words):
    len_text = len(words)
    masked_words = []
    for i in range(len_text - 1):
        masked_words.append(words[0:i] + ['[UNK]'] + words[i + 1:])
    # list of words
    return masked_words


def get_important_scores(words, pred_score, pred_indices, victim_model, config, max_len):
    masked_words = _get_masked(words)
    # List of text of masked words for each function snippet.
    # Mask one word at a time for each function.
    # There are as many [MASK]ed sentences as the no. of words.
    masked_codes = [' '.join(words) for words in masked_words]
    all_masked_code_ids = []
    all_attention_masks = []
    all_segs = []
    for code in masked_codes:
        inputs = victim_model['tokenizer'].encode_plus(code, None, add_special_tokens=True, max_length=config['max_source_length'])
        input_ids = inputs["input_ids"]
        attention_mask = [1] * len(input_ids)

        padding_length = max_len - len(input_ids)
        input_ids = input_ids + (padding_length * [0])
        attention_mask = attention_mask + (padding_length * [0])

        all_masked_code_ids.append(torch.tensor(input_ids))
        all_attention_masks.append(torch.tensor(attention_mask))

    all_masked_codes =  torch.stack(all_masked_code_ids).to(config['device'])
    all_attention_masks = torch.stack(all_attention_masks).to(config['device'])

    eval_data = TensorDataset(all_masked_codes)
    eval_attention_masks = TensorDataset(all_attention_masks)

    # Run prediction for full data
    eval_sampler = SequentialSampler(eval_data)
    eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=config['batch_size'])

    eval_attention_sampler = SequentialSampler(eval_attention_masks)
    eval_attention_dataloader = DataLoader(eval_attention_masks, sampler=eval_attention_sampler, batch_size=config['batch_size'])

    leave_1_probs = []
    for masked_input, masked_attention in zip(eval_dataloader, eval_attention_dataloader):
        masked_input_batch = masked_input[0]
        masked_attention_batch = masked_attention[0]

        if config['victim_model'] == 'codet5':
            leave_1_prob_batch = victim_model['model'].generate(masked_input_batch,
                                                                output_scores=True,
                                                                return_dict_in_generate=True,
                                                                early_stopping=config['task']=='summarization',
                                                                max_length=config['max_target_length'])
            scores = leave_1_prob_batch.scores

            # Reshape the scores: B x len(pred_seq) x vocab_size
            batch_scores_stack = torch.stack(list(scores))
            pred_len, batch_size, vocab = batch_scores_stack.shape
            batch_scores_stack = batch_scores_stack.reshape(batch_size, pred_len, vocab)

        elif config['victim_model'] == 'codebert' or config['victim_model'] == 'roberta':
            preds, scores = victim_model['model'](source_ids=masked_input_batch,
                                                  source_mask=masked_attention_batch)

            # Scores: B x len(pred_seq) x vocab_size
            batch_scores_stack = torch.stack(list(scores))

        # Do not use the index of the original predicted sequence
        # Calculate the max scores based on the highest values of the current predicted sequence
        if config['use_pred_idx'] == 0:
            max_scores_wo_pred_idx, max_scores_pred_idx = torch.max(batch_scores_stack, dim=2)
            max_scores = torch.sum(max_scores_wo_pred_idx, dim=1)

        # Use the index of the original predicted sequence
        # Calculate the max scores based on the current values at the index of the original predicted sequence
        # Original Pred Sequence Text: "<s> Two sum </s>"; Orginal pred Indices: [1, 100, 200, 2]
        # Calculate the scores only at these indices
        elif config['use_pred_idx'] == 1:
            max_scores_with_idx = torch.index_select(batch_scores_stack, 2, torch.tensor(pred_indices).to(config['device']))
            max_scores_idx_batch = torch.diagonal(max_scores_with_idx, dim1=-2, dim2=-1)
            max_scores = torch.sum(max_scores_idx_batch, dim=1)

        leave_1_probs.extend(max_scores)

    leave_1_probs = torch.stack(leave_1_probs)  # num_words, 1

    import_scores = pred_score - leave_1_probs
    # Index of the most important words and their scores in the I/P code
    index_important_words = sorted(enumerate(import_scores), key=lambda x: x[1], reverse=True)
    return index_important_words


def get_important_scores_graphcodebert(words, pred_score, pred_indices, victim_model, config, max_len):
    masked_words = _get_masked(words)
    # List of text of masked words for each function snippet.
    # Mask one word at a time for each function.
    # There are as many [MASK]ed sentences as the no. of words.
    masked_codes = [' '.join(words) for words in masked_words]
    all_masked_code_ids = []
    all_source_mask_ids = []
    all_attention_masks = []
    all_position_idx = []
    all_segs = []
    for code in masked_codes:
        examples = [code]
        input_ids, code_mask, position_idx, attention_mask = convert_examples_to_features(examples, victim_model['tokenizer'], config)
        # input_ids = input_ids.unsqueeze(0)
        # code_mask = code_mask.unsqueeze(0)
        # position_idx = position_idx.unsqueeze(0)

        all_masked_code_ids.append(input_ids)
        all_source_mask_ids.append(code_mask)
        all_position_idx.append(position_idx)
        all_attention_masks.append(attention_mask)

    all_masked_codes =  torch.stack(all_masked_code_ids).to(config['device'])
    all_source_masks = torch.stack(all_source_mask_ids).to(config['device'])
    all_position_idx = torch.stack(all_position_idx).to(config['device'])
    all_attention_masks = torch.stack(all_attention_masks).to(config['device'])

    eval_data = TensorDataset(all_masked_codes)
    eval_source_masks = TensorDataset(all_source_masks)
    eval_position_idx = TensorDataset(all_position_idx)
    eval_attention_masks = TensorDataset(all_attention_masks)

    # Run prediction for full data
    eval_sampler = SequentialSampler(eval_data)
    eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=config['batch_size'])

    eval_sm_sampler = SequentialSampler(eval_source_masks)
    eval_sm_dataloader = DataLoader(eval_source_masks, sampler=eval_sm_sampler, batch_size=config['batch_size'])

    eval_position_sampler = SequentialSampler(eval_position_idx)
    eval_position_dataloader = DataLoader(eval_position_idx, sampler=eval_position_sampler, batch_size=config['batch_size'])

    eval_attention_sampler = SequentialSampler(eval_attention_masks)
    eval_attention_dataloader = DataLoader(eval_attention_masks, sampler=eval_attention_sampler, batch_size=config['batch_size'])

    leave_1_probs = []
    for masked_input, masked_sm, masked_pos, masked_attention in zip(eval_dataloader, eval_sm_dataloader, eval_position_dataloader, eval_attention_dataloader):
        masked_input_batch = masked_input[0]
        masked_sm_batch = masked_sm[0]
        masked_pos_batch = masked_pos[0]
        masked_attention_batch = masked_attention[0]

        preds, scores = victim_model['model'](masked_input_batch,
                                              masked_sm_batch,
                                              masked_pos_batch,
                                              masked_attention_batch)

        # Scores: B x len(pred_seq) x vocab_size
        batch_scores_stack = torch.stack(list(scores))

        # Do not use the index of the original predicted sequence
        # Calculate the max scores based on the highest values of the current predicted sequence
        if config['use_pred_idx'] == 0:
            max_scores_wo_pred_idx, max_scores_pred_idx = torch.max(batch_scores_stack, dim=2)
            max_scores = torch.sum(max_scores_wo_pred_idx, dim=1)

        # Use the index of the original predicted sequence
        # Calculate the max scores based on the current values at the index of the original predicted sequence
        # Original Pred Sequence Text: "<s> Two sum </s>"; Orginal pred Indices: [1, 100, 200, 2]
        # Calculate the scores only at these indices
        elif config['use_pred_idx'] == 1:
            max_scores_with_idx = torch.index_select(batch_scores_stack, 2, torch.tensor(pred_indices).to(config['device']))
            max_scores_idx_batch = torch.diagonal(max_scores_with_idx, dim1=-2, dim2=-1)
            max_scores = torch.sum(max_scores_idx_batch, dim=1)

        leave_1_probs.extend(max_scores)

    leave_1_probs = torch.stack(leave_1_probs)  # num_words, 1

    import_scores = pred_score - leave_1_probs
    # Index of the most important words and their scores in the I/P code
    index_important_words = sorted(enumerate(import_scores), key=lambda x: x[1], reverse=True)
    return index_important_words