fed_train.py
9.09 KB
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
217
218
219
220
221
222
import utils
import copy
from collections import OrderedDict
import model
import dataset
import importlib
importlib.reload(utils)
importlib.reload(model)
importlib.reload(dataset)
from utils import *
import torch.quantization
def add_args(parser):
# parser.add_argument('--model', type=str, default='moderate-cnn',
# help='neural network used in training')
parser.add_argument('--dataset', type=str, default='cifar10', metavar='N',
help='dataset used for training')
parser.add_argument('--fold_num', type=int, default=0,
help='5-fold, 0 ~ 4')
parser.add_argument('--batch_size', type=int, default=256, metavar='N',
help='input batch size for training')
parser.add_argument('--lr', type=float, default=0.002, metavar='LR',
help='learning rate')
parser.add_argument('--n_nets', type=int, default=100, metavar='NN',
help='number of workers in a distributed cluster')
parser.add_argument('--comm_type', type=str, default='fedtwa',
help='which type of communication strategy is going to be used: layerwise/blockwise')
parser.add_argument('--comm_round', type=int, default=10,
help='how many round of communications we shoud use')
args = parser.parse_args(args=[])
return args
def start_fedavg(fed_model, args,
train_data_set,
data_idx_map,
net_data_count,
testloader,
device):
print("start fed avg")
criterion = nn.CrossEntropyLoss()
C = 0.1
num_edge = int(max(C * args.n_nets, 1))
total_data_count = 0
for _, data_count in net_data_count.items():
total_data_count += data_count
print("total data: %d" % total_data_count)
# quantize
# fed_model.eval()
# torch.jit.save(torch.jit.script(fed_model), './float.pth')
# return
fed_model.fuse_model()
# modules_to_fuse = [['conv1', 'relu1'], ['conv2', 'relu2'], ['conv3', 'relu3']]
# torch.quantization.fuse_modules(fed_model, modules_to_fuse, inplace=True)
fed_model.qconfig = torch.quantization.get_default_qat_qconfig('fbgemm')
torch.quantization.prepare_qat(fed_model, inplace=True)
# for making shape of weight_fake_quant.scale
train_data_set.set_idx_map([0])
fed_model(torch.from_numpy(np.expand_dims(train_data_set[0][0], axis=0)).float())
edges, _, _ = init_models(args.n_nets, args)
# edges = [copy.deepcopy(fed_model) for net_cnt in range(args.n_nets)]
for edge_now in edges:
edge_now.fuse_model()
edge_now.qconfig = torch.quantization.get_default_qat_qconfig('fbgemm')
torch.quantization.prepare_qat(edge_now, inplace=True)
edge_now(torch.from_numpy(np.expand_dims(train_data_set[0][0], axis=0)).float())
# print('quantized \n', edges[edge_index].conv1)
# end
for cr in range(1, args.comm_round + 1):
print("Communication round : %d" % (cr))
np.random.seed(cr) # make sure for each comparison, select the same clients each round
selected_edge = np.random.choice(args.n_nets, num_edge, replace=False)
print("selected edge", selected_edge)
for edge_progress, edge_index in enumerate(selected_edge):
train_data_set.set_idx_map(data_idx_map[edge_index])
train_loader = torch.utils.data.DataLoader(train_data_set, batch_size=args.batch_size,
shuffle=True, num_workers=2)
print("[%2d/%2d] edge: %d, data len: %d" % (edge_progress, len(selected_edge), edge_index, len(train_data_set)))
edges[edge_index] = copy.deepcopy(fed_model)
edges[edge_index].to(device)
edges[edge_index].train()
edge_opt = optim.Adam(params=edges[edge_index].parameters(), lr=args.lr)
# train
for data_idx, (inputs, labels) in enumerate(train_loader):
inputs, labels = inputs.float().to(device), labels.long().to(device)
edge_opt.zero_grad()
# edge_opt[edge_index].zero_grad()
edge_pred = edges[edge_index](inputs)
edge_loss = criterion(edge_pred, labels)
edge_loss.backward()
edge_opt.step()
# edge_opt[edge_index].step()
edge_loss = edge_loss.item()
if data_idx % 100 == 0:
print('[%4d] loss: %.3f' % (data_idx, edge_loss))
break
edges[edge_index].to('cpu')
# print(edge_index)
# local_state = edges[edge_index].state_dict()
# for key in edges[edge_index].state_dict().keys():
# if 'activation_post_process' in key or 'fake_quant' in key:
# print(key, local_state[key])
# print()
# return
# cal weight using fed avg
update_state = OrderedDict()
for k, edge in enumerate(edges):
local_state = edge.state_dict()
for key in fed_model.state_dict().keys():
# if 'zero_point' in key:
# print(local_state[key])
if 'activation_post_process' in key or 'fake_quant' in key:
if k == 0:
update_state[key] = local_state[key]
else:
update_state[key] += local_state[key]
elif 'enable' in key:
update_state[key] = local_state[key]
else:
if k == 0:
update_state[key] = local_state[key] * (net_data_count[k] / total_data_count)
else:
update_state[key] += local_state[key] * (net_data_count[k] / total_data_count)
# break
for key in update_state.keys():
if 'enable' in key:
continue
if 'activation_post_process' in key or 'fake_quant' in key:
# print(key, update_state[key], update_state[key].type())
# print(key, update_state[key])
if torch.is_floating_point(update_state[key]):
update_state[key] = update_state[key] / args.n_nets
else:
update_state[key] = torch.floor_divide(update_state[key], args.n_nets)
# print(update_state[key])
fed_model.load_state_dict(update_state)
if cr % 1 == 0:
fed_model.to(device)
fed_model.eval()
total_loss = 0.0
cnt = 0
step_acc = 0.0
with torch.no_grad():
for i, data in enumerate(testloader):
inputs, labels = data
inputs, labels = inputs.float().to(device), labels.long().to(device)
outputs = fed_model(inputs)
_, preds = torch.max(outputs, 1)
loss = criterion(outputs, labels)
cnt += inputs.shape[0]
corr_sum = torch.sum(preds == labels.data)
step_acc += corr_sum.double()
running_loss = loss.item() * inputs.shape[0]
total_loss += running_loss
if i % 200 == 0:
print('test [%4d] loss: %.3f' % (i, loss.item()))
break
print((step_acc / cnt).item())
print(total_loss / cnt)
fed_model.to('cpu')
quantized_fed_model = torch.quantization.convert(fed_model.eval(), inplace=False)
torch.jit.save(torch.jit.script(quantized_fed_model), './quan.pth')
def start_train():
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
args = add_args(argparse.ArgumentParser())
seed = 0
np.random.seed(seed)
torch.manual_seed(seed)
print("Loading data...")
# kwargs = {"./dataset/DoS_dataset.csv" : './DoS_dataset.txt',
# "./dataset/Fuzzy_dataset.csv" : './Fuzzy_dataset.txt',
# "./dataset/RPM_dataset.csv" : './RPM_dataset.txt',
# "./dataset/gear_dataset.csv" : './gear_dataset.txt'
# }
kwargs = {"./dataset/DoS_dataset.csv" : './DoS_dataset.txt'}
train_data_set, data_idx_map, net_class_count, net_data_count, test_data_set = dataset.GetCanDatasetUsingTxtKwarg(args.n_nets, args.fold_num, **kwargs)
testloader = torch.utils.data.DataLoader(test_data_set, batch_size=args.batch_size,
shuffle=False, num_workers=2)
run_benchmark('./quan.pth', testloader)
run_benchmark('./float.pth', testloader)
# run_benchmark('./quan.pth', testloader)
fed_model = model.Net()
args.comm_type = 'fedavg'
if args.comm_type == "fedavg":
start_fedavg(fed_model, args,
train_data_set,
data_idx_map,
net_data_count,
testloader,
device)
if __name__ == "__main__":
start_train()