Augmentation-Strategy-Optimization-for-Language-Understanding/attack.py at main · NYCU-MLLab/Augmentation-Strategy-Optimization-for-Language-Understanding · 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
import textattack
import transformers
import pandas as pd
from argparse import ArgumentParser


NUM_CLASSES = 2


class AugmentArgs:
    input_csv: str
    output_csv: str
    input_column: str
    overwrite: bool = True

    @classmethod
    def _add_parser_args(cls, parser):
        parser.add_argument(
            "--datapath-prefix",
            required=True,
            type=str,
            help="Path of input dataset to attack.",
        )

        parser.add_argument(
            "--attack-method",
            required=True,
            type=str,
            help="Name of attack recipe."
        )

        parser.add_argument(
            "--target-model",
            required=False,
            type=str,
            default="roberta-base"
        )


def get_dataset(
    datapath: str
):
    input_df = pd.read_csv(datapath)
    sentences = input_df["sentence"].to_list()
    sentences = [str(sentence) for sentence in sentences]
    labels = input_df["label"].to_list()
    pre_dataset = list(zip(sentences, labels))

    global NUM_CLASSES
    NUM_CLASSES = len(set(labels))

    return textattack.datasets.Dataset(pre_dataset)


def get_attack_module(
    attack_method_name: str,
    model_wrapper
):
    if attack_method_name == "pwws":
        attack = textattack.attack_recipes.PWWSRen2019.build(model_wrapper)
    elif attack_method_name == "fast-alzantot":
        attack = textattack.attack_recipes.FasterGeneticAlgorithmJia2019.build(model_wrapper)
    elif attack_method_name == "iga":
        attack = textattack.attack_recipes.IGAWang2019.build(model_wrapper)
    elif attack_method_name == "textfooler":
        attack = textattack.attack_recipes.TextFoolerJin2019.build(model_wrapper)
    elif attack_method_name == "hotflip":
        attack = textattack.attack_recipes.HotFlipEbrahimi2017.build(model_wrapper)
    elif attack_method_name == "bae":
        attack = textattack.attack_recipes.BAEGarg2019.build(model_wrapper)
    elif attack_method_name == "deepwordbug":
        attack = textattack.attack_recipes.DeepWordBugGao2018.build(model_wrapper)
    elif attack_method_name == "input-reduction":
        attack = textattack.attack_recipes.InputReductionFeng2018.build(model_wrapper)
    elif attack_method_name == "kuleshov":
        attack = textattack.attack_recipes.Kuleshov2017.build(model_wrapper)
    elif attack_method_name == "pso":
        attack = textattack.attack_recipes.PSOZang2020.build(model_wrapper)
    elif attack_method_name == "textbugger":
        attack = textattack.attack_recipes.TextBuggerLi2018.build(model_wrapper)
    else:
        raise ValueError("Unsupported Attack Method")

    return attack


def main():
    parser = ArgumentParser()
    AugmentArgs._add_parser_args(parser)
    args = parser.parse_args()

    train_ds = get_dataset(args.datapath_prefix + "_train.csv")
    test_ds = get_dataset(args.datapath_prefix + "_test.csv")
    model = transformers.AutoModelForSequenceClassification.from_pretrained(args.target_model, num_labels=NUM_CLASSES)
    tokenizer = transformers.AutoTokenizer.from_pretrained(args.target_model)
    model_wrapper = textattack.models.wrappers.HuggingFaceModelWrapper(model, tokenizer)

    attack = get_attack_module(args.attack_method, model_wrapper)

    training_args = textattack.TrainingArgs(
        num_epochs=9,
        num_clean_epochs=3,
        attack_epoch_interval=2,
        num_train_adv_examples=6000,
        learning_rate=1e-5,
        num_warmup_steps=0.06,
        attack_num_workers_per_device=9,
        per_device_train_batch_size=8,
        gradient_accumulation_steps=4,
        log_to_tb=True,
    )
    trainer = textattack.Trainer(
        model_wrapper,
        "classification",
        attack,
        train_ds,
        test_ds,
        training_args
    )

    trainer.train()

    noisy_ds = []
    noisy_names = ["stack_eda", "eda", "embedding", "clare", "checklist", "char", "backtrans_de", "backtrans_ru", "backtrans_zh", "spell"]
    noisy_ds.append(get_dataset(args.datapath_prefix + "_stack_eda.csv"))
    noisy_ds.append(get_dataset(args.datapath_prefix + "_eda.csv"))
    noisy_ds.append(get_dataset(args.datapath_prefix + "_embedding.csv"))
    noisy_ds.append(get_dataset(args.datapath_prefix + "_clare.csv"))
    noisy_ds.append(get_dataset(args.datapath_prefix + "_checklist.csv"))
    noisy_ds.append(get_dataset(args.datapath_prefix + "_char.csv"))
    noisy_ds.append(get_dataset(args.datapath_prefix + "_backtrans_de.csv"))
    noisy_ds.append(get_dataset(args.datapath_prefix + "_backtrans_ru.csv"))
    noisy_ds.append(get_dataset(args.datapath_prefix + "_backtrans_zh.csv"))
    noisy_ds.append(get_dataset(args.datapath_prefix + "_spell.csv"))

    for idx, nds in enumerate(noisy_ds):
        trainer = textattack.Trainer(
            model_wrapper,
            "classification",
            attack,
            train_ds,
            nds,
            training_args
        )
        print(noisy_names[idx])
        trainer.evaluate()


if __name__ == '__main__':
    main()