flexfl/main.py at main · IBM/flexfl · 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
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
from torch.utils.data import DataLoader
from config import *
from dataset.dataset import *
from statistic.collect_stat import CollectStatistics
from util.util import split_data, NodeSampler
import numpy as np
import random
from model.model import Model
from util.util import DatasetSplit

import compressed_update
import partial_participation

if device.type != 'cpu':
    torch.cuda.set_device(device)


if __name__ == "__main__":
    stat = CollectStatistics(results_eval_file_prefix=results_file_prefix)

    for seed in simulations:

        random.seed(seed)
        np.random.seed(seed)  # numpy
        torch.manual_seed(seed)  # cpu
        torch.cuda.manual_seed(seed)  # gpu
        torch.backends.cudnn.deterministic = True  # cudnn

        data_train, data_test = load_data(dataset, dataset_file_path, 'cpu')
        data_train_loader = DataLoader(data_train, batch_size=batch_size_eval, shuffle=True, num_workers=0)
        data_test_loader = DataLoader(data_test, batch_size=batch_size_eval, num_workers=0)
        dict_users = split_data(dataset, data_train, n_nodes)
        if n_nodes is None:
            n_nodes = len(dict_users)

        node_sampler = NodeSampler(n_nodes, permutation=use_permute)

        model = Model(seed, step_size, model_name=model_name, device=device, flatten_weight=True,
                      pretrained_model_file=load_model_file)

        train_loader_list = []
        dataiter_list = []
        for n in range(n_nodes):
            train_loader_list.append(
                DataLoader(DatasetSplit(data_train, dict_users[n]), batch_size=batch_size_train, shuffle=True))
            dataiter_list.append(iter(train_loader_list[n]))


        def sample_minibatch(n):
            try:
                images, labels = dataiter_list[n].next()
                if len(images) < batch_size_train:
                    dataiter_list[n] = iter(train_loader_list[n])
                    images, labels = dataiter_list[n].next()
            except StopIteration:
                dataiter_list[n] = iter(train_loader_list[n])
                images, labels = dataiter_list[n].next()

            return images, labels

        def sample_full_batch(n):
            images = []
            labels = []
            for i in range(len(train_loader_list[n].dataset)):
                images.append(train_loader_list[n].dataset[i][0])

                l = train_loader_list[n].dataset[i][1]
                if not isinstance(l, torch.Tensor):
                    l = torch.as_tensor(l)
                labels.append(l)

            return torch.stack(images), torch.stack(labels)

        w_global = model.get_weight()   # Get initial weight

        num_iter = 0
        last_output = 0
        last_amplify = 0
        last_save_latest = 0
        last_save_checkpoint = 0

        compression_method_str = compression_adaptive_method.split('-')

        # Client compression configs
        w_residual_updates_at_node = []
        compressor_at_node = []
        sum_comm_cost_at_node = []
        count_comm_at_node = []

        for n in range(n_nodes):
            sum_comm_cost_at_node.append(0.0)
            count_comm_at_node.append(0)
            w_residual_updates_at_node.append(torch.zeros(w_global.shape[0]).to(device))  # TODO: Check whether to use to(device)

            if compression_method_str[0] == 'lyapunov':
                compressor_at_node.append(compressed_update.CompressedLyapunov(node=n, target_average_cost=target_avg_comm_cost, v=lyapunov_v, init_queue=lyapunov_init_queue))
            elif compression_method_str[0] == 'fixed':
                amount_of_transmission = float(compression_method_str[1])
                compressor_at_node.append(compressed_update.CompressedNoneOrFixedRandom(node=n, target_average_cost=target_avg_comm_cost,
                                                                                        amount_of_transmission=amount_of_transmission))

        # Server compression configs
        sum_comm_cost_at_server = 0.0
        count_comm_at_server = 0
        w_residual_updates_at_server = torch.zeros(w_global.shape[0]).to(device)
        compressor_at_server = None
        if compression_method_str[0] == 'lyapunov':
            compressor_at_server = compressed_update.CompressedLyapunov(node=-1, target_average_cost=target_avg_comm_cost, v=lyapunov_v, init_queue=lyapunov_init_queue)
        elif compression_method_str[0] == 'fixed':
            amount_of_transmission = float(compression_method_str[1])
            compressor_at_server = compressed_update.CompressedNoneOrFixedRandom(node=-1, target_average_cost=target_avg_comm_cost,
                                                                                 amount_of_transmission=amount_of_transmission)

        sum_part_cost_at_node = []
        sum_obj_cost_at_node = []
        count_part_at_node = []
        part_handler_at_node = []
        for n in range(n_nodes):
            sum_part_cost_at_node.append(0.0)
            sum_obj_cost_at_node.append(0.0)
            count_part_at_node.append(0)
            if compression_method_str[0] == 'lyapunov':
                part_handler_at_node.append(partial_participation.ParticipationLyapunov(n, target_avg_participation_cost, v=lyapunov_v, init_queue=lyapunov_init_queue))
            elif compression_method_str[0] == 'fixed':
                part_handler_at_node.append(partial_participation.ParticipationStatic(n, target_avg_participation_cost))

        while True:
            print('seed', seed, '  iteration', num_iter)

            accumulated = 0

            for n in range(n_nodes):
                participation_prob = part_handler_at_node[n].get_participation(num_iter)
                if np.random.binomial(1, participation_prob) == 1:
                    model.assign_weight(w_global)
                    model.model.train()

                    for i in range(0, iters_per_round):
                        images, labels = sample_minibatch(n)

                        images, labels = images.to(device), labels.to(device)

                        if transform_train is not None:
                            images = transform_train(images)

                        model.optimizer.zero_grad()
                        output = model.model(images)
                        loss = model.loss_fn(output, labels)
                        loss.backward()
                        model.optimizer.step()

                    w_tmp = model.get_weight()  # deepcopy is already included here
                    w_tmp -= w_global  # This is the difference (i.e., update) in this round
                    w_tmp /= participation_prob
                    sum_part_cost_at_node[n] += partial_participation.participation_cost_at_node(n, num_iter)
                else:
                    w_tmp = None

                sum_obj_cost_at_node[n] += 1.0 / participation_prob
                count_part_at_node[n] += 1
                stat.collect_stat_part_cost(seed, num_iter, n, count_part_at_node[n],
                                            sum_obj_cost_at_node[n] / count_part_at_node[n],
                                            sum_part_cost_at_node[n] / count_part_at_node[n])


                w_tmp, w_residual_updates_at_node[n] = compressor_at_node[n].get_transmitted_and_residual(num_iter, w_tmp, w_residual_updates_at_node[n])

                cost_instantaneous = compressed_update.transmission_cost_at_node(n, num_iter, w_tmp.shape[0], torch.count_nonzero(w_tmp))
                sum_comm_cost_at_node[n] += cost_instantaneous
                count_comm_at_node[n] += 1
                stat.collect_stat_comm_cost(seed, num_iter, n, count_comm_at_node[n],
                                            torch.count_nonzero(w_tmp).item(), cost_instantaneous,
                                            sum_comm_cost_at_node[n] / count_comm_at_node[n])

                if accumulated == 0:  # accumulated weights
                    w_accumulate = w_tmp
                    # Note: w_tmp cannot be used after this
                else:
                    w_accumulate += w_tmp

                accumulated += 1

            if accumulated > 0:
                w_tmp = torch.div(w_accumulate, torch.tensor(accumulated).to(device)).view(-1)
            else:
                w_tmp = torch.zeros(w_global.shape[0]).to(device)

            w_tmp, w_residual_updates_at_server = compressor_at_server.get_transmitted_and_residual(num_iter, w_tmp, w_residual_updates_at_server)

            cost_instantaneous = compressed_update.transmission_cost_at_node(-1, num_iter, w_tmp.shape[0], torch.count_nonzero(w_tmp))
            sum_comm_cost_at_server += cost_instantaneous
            count_comm_at_server += 1
            stat.collect_stat_comm_cost(seed, num_iter, -1, count_comm_at_server,
                                        torch.count_nonzero(w_tmp).item(), cost_instantaneous,
                                        sum_comm_cost_at_server / count_comm_at_server)

            w_global += w_tmp

            num_iter = num_iter + iters_per_round

            if save_checkpoint and num_iter - last_save_checkpoint >= iters_checkpoint:
                torch.save(model.model.state_dict(), save_model_file + '-checkpoint-sim-' + str(seed) + '-iter-' + str(num_iter))
                last_save_checkpoint = num_iter

            if num_iter - last_output >= min_iters_per_eval:
                stat.collect_stat_eval(seed, num_iter, model, data_train_loader, data_test_loader, w_global)
                last_output = num_iter

            if num_iter >= max_iter:
                break

        del model
        del w_global
        del w_accumulate

        torch.cuda.empty_cache()