CLUE/main_viz.py at main · Restuccia-Group/CLUE · 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
189
190
191
192
import argparse
import os
import pdb
import pickle
import random
import shutil
import time
import copy
from copy import deepcopy
from collections import OrderedDict

import arg_parser
import matplotlib.pyplot as plt
import numpy as np
import torch
import torch.backends.cudnn as cudnn
import torch.nn as nn
import torch.nn.functional as F
import torch.optim
import torch.utils.data
import torchvision.datasets as datasets
import torchvision.models as models
import torchvision.transforms as transforms
from torch.utils.data.sampler import SubsetRandomSampler
from trainer import train, validate
import utils
import unlearn

from viz_utils.tsne import extract_features, get_tsne
best_sa = 0

def reduce_data(data_set, percentage, seed):
    valid_idx = []
    rng = np.random.RandomState(seed)
    for i in range(max(data_set.targets) + 1):
        class_idx = np.where(data_set.targets == i)[0]
        valid_idx.append(
            rng.choice(class_idx, int(percentage * len(class_idx)), replace=False)
        )
    valid_idx = np.hstack(valid_idx)
    train_set_copy = copy.deepcopy(data_set)

    data_set.data = train_set_copy.data[valid_idx]
    data_set.targets = train_set_copy.targets[valid_idx]
    return data_set

def main():
    args = arg_parser.parse_args()

    if torch.cuda.is_available():
        torch.cuda.set_device(int(args.gpu))
        device = torch.device(f"cuda:{int(args.gpu)}")
    else:
        device = torch.device("cpu")

    os.makedirs(args.save_dir, exist_ok=True)
    if args.seed:
        utils.setup_seed(args.seed)
    seed = args.seed
    # prepare dataset
    (
        model,
        train_loader_full,
        val_loader,
        test_loader,
        marked_loader,
    ) = utils.setup_model_dataset(args)
    model.cuda()

    def replace_loader_dataset(
        dataset, batch_size=args.batch_size, seed=1, shuffle=True
    ):
        utils.setup_seed(seed)
        return torch.utils.data.DataLoader(
            dataset,
            batch_size=batch_size,
            num_workers=0,
            pin_memory=True,
            shuffle=shuffle,
        )

    forget_dataset = copy.deepcopy(marked_loader.dataset)
    if args.dataset == "svhn":
        try:
            marked = forget_dataset.targets < 0
        except:
            marked = forget_dataset.labels < 0
        forget_dataset.data = forget_dataset.data[marked]
        try:
            forget_dataset.targets = -forget_dataset.targets[marked] - 1
        except:
            forget_dataset.labels = -forget_dataset.labels[marked] - 1
        forget_loader = replace_loader_dataset(forget_dataset, seed=seed, shuffle=True)
        retain_dataset = copy.deepcopy(marked_loader.dataset)
        try:
            marked = retain_dataset.targets >= 0
        except:
            marked = retain_dataset.labels >= 0
        retain_dataset.data = retain_dataset.data[marked]
        try:
            retain_dataset.targets = retain_dataset.targets[marked]
        except:
            retain_dataset.labels = retain_dataset.labels[marked]
        retain_dataset = reduce_data(retain_dataset, args.retain_percentage, seed)  #This will reduce the size of the retain set availabel to the unlearning algorithm
        retain_loader = replace_loader_dataset(retain_dataset, seed=seed, shuffle=True)
        assert len(forget_dataset) + args.retain_percentage * len(retain_dataset) == len(
            train_loader_full.dataset
        )
    else:
        try:
            marked = forget_dataset.targets < 0
            forget_dataset.data = forget_dataset.data[marked]
            forget_dataset.targets = -forget_dataset.targets[marked] - 1
            forget_loader = replace_loader_dataset(
                forget_dataset, seed=seed, shuffle=True
            )
            retain_dataset = copy.deepcopy(marked_loader.dataset)
            marked = retain_dataset.targets >= 0
            retain_dataset.data = retain_dataset.data[marked]
            retain_dataset.targets = retain_dataset.targets[marked]
            retain_dataset = reduce_data(retain_dataset, args.retain_percentage, seed)  #This will reduce the size of the retain set availabel to the unlearning algorithm
            retain_loader = replace_loader_dataset(
                retain_dataset, seed=seed, shuffle=True
            )
            # assert len(forget_dataset) + args.retain_percentage * len(retain_dataset) == len(
            #     train_loader_full.dataset
            # )
        except:
            marked = forget_dataset.targets < 0
            forget_dataset.imgs = forget_dataset.imgs[marked]
            forget_dataset.targets = -forget_dataset.targets[marked] - 1
            forget_loader = replace_loader_dataset(
                forget_dataset, seed=seed, shuffle=True
            )
            retain_dataset = copy.deepcopy(marked_loader.dataset)
            marked = retain_dataset.targets >= 0
            retain_dataset.imgs = retain_dataset.imgs[marked]
            retain_dataset.targets = retain_dataset.targets[marked]
            retain_dataset = reduce_data(retain_dataset, args.retain_percentage, seed)  #This will reduce the size of the retain set availabel to the unlearning algorithm
            retain_loader = replace_loader_dataset(
                retain_dataset, seed=seed, shuffle=True
            )
            assert len(forget_dataset) + args.retain_percentage * len(retain_dataset) == len(
                train_loader_full.dataset
            )

    print(f"number of retain dataset {len(retain_dataset)}")
    print(f"number of forget dataset {len(forget_dataset)}")
    unlearn_data_loaders = OrderedDict(
        retain=retain_loader, forget=forget_loader, val=val_loader, test=test_loader
    )

    criterion = nn.CrossEntropyLoss()
    evaluation_result = None

    if args.resume:
        checkpoint = unlearn.load_unlearn_checkpoint(model, device, args)

    if args.resume and checkpoint is not None:
        model, evaluation_result = checkpoint
    else:
        if not args.chenyaofo:
            checkpoint = torch.load(args.model_path, map_location=device)
            if "state_dict" in checkpoint.keys():
                checkpoint = checkpoint["state_dict"]

            if args.unlearn != "retrain":
                model.load_state_dict(checkpoint, strict=False)


    # features, labels = extract_features(model, train_loader_full)
    # print(f"The shape of features and labels are : {features.shape}, {labels.shape}")

    #Inspecting the ViT

    # for name, module in model.named_modules():
    #     print(name)
    #     print(module)

    #     print("\n")

    # raise Exception

    X,Y = get_tsne(model, train_loader_full)
    print(f"The shape of features and labels are : {X.shape}, {Y.shape}")

    np.savetxt(f"./viz_assets/{args.arch}_{args.dataset}_data.csv", X)
    np.savetxt(f"./viz_assets/{args.arch}_{args.dataset}_labels.csv", Y)


if __name__ == "__main__":
    main()