FedTHE/BRFL/run.py at master · LINs-lab/FedTHE · 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
# -*- coding: utf-8 -*-
import argparse
import os
import torch
import numpy as np
from torch.utils.data import Subset

import prepare_data as prepare_data
import partition_data as partition_data
import create_ood_test as create_ood_test

if __name__ == "__main__":
    # Hyperparameters.
    parser = argparse.ArgumentParser()
    parser.add_argument('--seed', type=int, default=6)
    parser.add_argument('--data_path', type=str, default=os.getcwd())
    parser.add_argument('--num_clients', type=int, default=20)
    parser.add_argument('--data_type', type=str, default="cifar10")
    parser.add_argument('--local_tr_ratio', type=float, default=0.6)
    parser.add_argument('--local_te_ratio', type=float, default=0.2)
    parser.add_argument('--non_iid_alpha', type=float, default=0.1)
    parser.add_argument(
        '--corr_severity',
        help="Severity of corruption (only affect corrupted tests), between 0 and 5",
        type=int,
        default=5
    )
    parser.add_argument(
        '--weighted_sampling_mixed_test',
        help="Whether sampling each test distribution approx. equally (for cifar10)",
        type=bool,
        default=True
    )

    args = parser.parse_args()

    random_state = np.random.RandomState(args.seed)
    # Sanity check.
    assert args.local_tr_ratio > 0 and args.local_te_ratio >= 0
    assert args.data_type == "cifar10" or args.data_type == "imagenet"

    dataset = {}
    if args.data_type == "cifar10":

        # Prepare dataset.
        dataset["cifar10_tr"] = prepare_data.get_dataset(
            data_name="cifar10", datasets_path=args.data_path, split="train",
        )
        dataset["cifar10_te"] = prepare_data.get_dataset(
            data_name="cifar10", datasets_path=args.data_path, split="test",
        )

        # Merge, then split.
        fl_data = torch.utils.data.ConcatDataset([dataset["cifar10_tr"], dataset["cifar10_te"]])
        fl_data.indices = list(range(len(dataset["cifar10_tr"]) + len(dataset["cifar10_te"])))
        fl_data.targets = dataset["cifar10_tr"].targets + dataset["cifar10_te"].targets

        non_iid_indices = partition_data.inter_client_non_iid_partition(
            fl_data, args.non_iid_alpha, random_state, args.num_clients
        )

        # Intra-split into local train/val/test.
        indices_per_client = partition_data.intra_client_uniform_partition(
            non_iid_indices, random_state, args.local_tr_ratio, args.local_te_ratio
        )

        # Create ID/OOD train & test datasets.
        fl_data_per_client = {}
        fl_data_per_client["train"], fl_data_per_client["val"], fl_data_per_client["test"] = {}, {}, {}
        for i, indices in indices_per_client.items():
            fl_data_per_client["train"][i] = Subset(fl_data, indices["train"])
            fl_data_per_client["val"][i] = Subset(fl_data, indices["val"])
            fl_data_per_client["test"][i] = Subset(fl_data, indices["test"])
        fl_data_per_client["corr_test"] = create_ood_test.get_corr_data(
            fl_data,
            indices_per_client,
            random_state,
            severity=args.corr_severity
        )  # common corruptions
        fl_data_per_client["ooc_test"] = create_ood_test.get_ooc_data(
            fl_data,
            indices_per_client,
            random_state,
        )  # out-of-client (label shift) test
        fl_data_per_client["natural_shift_test"] = create_ood_test.get_natural_shift_data(
            fl_data,
            indices_per_client,
            random_state,
            data_path=args.data_path,
            data_name="cifar10_1",
        )  # cifar10.1 (natural shift) test
        fl_data_per_client["mixed_test"] = create_ood_test.get_mixed_data(
            fl_data_per_client,
            random_state,
            weighted_sampling=args.weighted_sampling_mixed_test,
        )  # mixed test

        # `fl_data_per_client` thus becomes a dict where fl_data_per_client[test_type][client_id] gives the corresponding local dataset.
        # Final operations, e.g., create dataloaders.
        test_loader = torch.utils.data.DataLoader(
            fl_data_per_client["test"][0],
            batch_size=32,
            shuffle=True,
            drop_last=False,
        )
        for imgs, labels in test_loader:
            imgs, labels = prepare_data.transform_data_batch(imgs, labels, is_training=False)

    elif args.data_type == "imagenet":

        # Prepare dataset.
        dataset["imagenet_tr"] = prepare_data.get_dataset(
            data_name="imagenet32", datasets_path=args.data_path, split="train",
        )
        dataset["imagenet_te"] = prepare_data.get_dataset(
            data_name="imagenet32", datasets_path=args.data_path, split="test",
        )

        # Merge, then split.
        fl_data = torch.utils.data.ConcatDataset([dataset["imagenet_tr"], dataset["imagenet_te"]])
        fl_data.indices = list(range(len(dataset["imagenet_tr"]) + len(dataset["imagenet_te"])))
        fl_data.targets = dataset["imagenet_tr"].targets + dataset["imagenet_te"].targets

        non_iid_indices = partition_data.inter_client_non_iid_partition(
            fl_data, args.non_iid_alpha, random_state, args.num_clients
        )

        # Intra-split into local train/val/test.
        indices_per_client = partition_data.intra_client_uniform_partition(
            non_iid_indices, random_state, args.local_tr_ratio, args.local_te_ratio
        )

        # Create ID/OOD train & test datasets.
        fl_data_per_client = {}
        fl_data_per_client["train"], fl_data_per_client["val"], fl_data_per_client["test"] = {}, {}, {}
        for i, indices in indices_per_client.items():
            fl_data_per_client["train"][i] = Subset(fl_data, indices["train"])
            fl_data_per_client["val"][i] = Subset(fl_data, indices["val"])
            fl_data_per_client["test"][i] = Subset(fl_data, indices["test"])
        fl_data_per_client["a_test"] = create_ood_test.get_natural_shift_data(
            fl_data,
            indices_per_client,
            random_state,
            data_path=args.data_path,
            data_name="imagenet_a",
        )  # ImageNet-A
        fl_data_per_client["v2_test"] = create_ood_test.get_natural_shift_data(
            fl_data,
            indices_per_client,
            random_state,
            data_path=args.data_path,
            data_name="imagenet_v2_matched-frequency",
        )  # ImageNet-V2
        fl_data_per_client["r_test"] = create_ood_test.get_natural_shift_data(
            fl_data,
            indices_per_client,
            random_state,
            data_path=args.data_path,
            data_name="imagenet_r",
        )  # ImageNet-R
        fl_data_per_client["mixed_test"] = create_ood_test.get_mixed_data(
            fl_data_per_client,
            random_state,
        )  # mixed test

        # `fl_data_per_client` thus becomes a dict where fl_data_per_client[test_type][client_id] gives the corresponding local dataset.
        # Final operations, e.g., create dataloaders.
        test_loader = torch.utils.data.DataLoader(
            fl_data_per_client["test"][0],
            batch_size=128,
            shuffle=True,
            drop_last=False,
        )
        for imgs, labels in test_loader:
            imgs, labels = prepare_data.transform_data_batch(imgs, labels, is_training=False)