-
Notifications
You must be signed in to change notification settings - Fork 2
/
zo_rge_main.py
188 lines (165 loc) · 7.12 KB
/
zo_rge_main.py
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
from os import path
from typing import Any
import torch
import torch.nn as nn
from tensorboardX import SummaryWriter
from tqdm import tqdm
from cezo_fl.coordinate_gradient_estimator import CoordinateGradientEstimator as CGE
from cezo_fl.models.cnn_fashion import CNN_FMNIST
from cezo_fl.models.cnn_mnist import CNN_MNIST
from cezo_fl.models.lenet import LeNet
from cezo_fl.models.lstm import CharLSTM
from cezo_fl.random_gradient_estimator import RandomGradientEstimator as RGE
from cezo_fl.util import model_helpers
from cezo_fl.util.checkpoint import CheckPoint
from cezo_fl.util.metrics import Metric, accuracy
from config import get_args_str, get_params
from preprocess import preprocess, use_sparsity_dict
def prepare_settings(args, device):
if args.dataset == "mnist":
model = CNN_MNIST().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(
model_helpers.get_trainable_model_parameters(model),
lr=args.lr,
weight_decay=1e-5,
momentum=args.momentum,
)
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.8)
elif args.dataset == "cifar10":
model = LeNet().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(
model_helpers.get_trainable_model_parameters(model),
lr=args.lr,
weight_decay=5e-4,
momentum=args.momentum,
)
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[200], gamma=0.1)
elif args.dataset == "fashion":
model = CNN_FMNIST().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(
model_helpers.get_trainable_model_parameters(model),
lr=args.lr,
weight_decay=1e-5,
momentum=args.momentum,
)
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[200], gamma=0.1)
elif args.dataset == "shakespeare":
model = CharLSTM().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(
model_helpers.get_trainable_model_parameters(model),
lr=args.lr,
momentum=0.9,
weight_decay=5e-4,
)
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[200], gamma=0.1)
if args.grad_estimate_method in ["rge-central", "rge-forward"]:
method = args.grad_estimate_method[4:]
print(f"Using RGE {method}")
grad_estimator = RGE(
model,
parameters=model_helpers.get_trainable_model_parameters(model),
mu=args.mu,
num_pert=args.num_pert,
grad_estimate_method=method,
device=device,
)
elif args.grad_estimate_method in ["cge-forward"]:
print("Using CGE forward")
grad_estimator = CGE(
model,
mu=args.mu,
device=device,
)
else:
raise Exception(f"Grad estimate method {args.grad_estimate_method} not supported")
return model, criterion, optimizer, scheduler, grad_estimator
def get_warmup_lr(args: Any, current_epoch: int, current_iter: int, iters_per_epoch: int) -> float:
overall_iterations = args.warmup_epochs * iters_per_epoch + 1
current_iterations = current_epoch * iters_per_epoch + current_iter + 1
return args.lr * current_iterations / overall_iterations
def train_model(epoch: int) -> tuple[float, float]:
model.train()
train_loss = Metric("train loss")
train_accuracy = Metric("train accuracy")
iter_per_epoch = len(train_loader)
with tqdm(total=iter_per_epoch, desc="Training:") as t, torch.no_grad():
for iteration, (images, labels) in enumerate(train_loader):
if epoch < args.warmup_epochs:
warmup_lr = get_warmup_lr(args, epoch, iteration, iter_per_epoch)
for p in optimizer.param_groups:
p["lr"] = warmup_lr
if device != torch.device("cpu"):
images, labels = images.to(device), labels.to(device)
# update models
optimizer.zero_grad()
grad_estimator.compute_grad(images, labels, criterion, seed=iteration**2 + iteration)
optimizer.step()
pred = model(images)
train_loss.update(criterion(pred, labels))
train_accuracy.update(accuracy(pred, labels))
t.set_postfix({"Loss": train_loss.avg, "Accuracy": train_accuracy.avg})
t.update(1)
if epoch > args.warmup_epochs:
scheduler.step()
return train_loss.avg, train_accuracy.avg
def eval_model(epoch: int) -> tuple[float, float]:
model.eval()
eval_loss = Metric("Eval loss")
eval_accuracy = Metric("Eval accuracy")
with torch.no_grad():
for _, (images, labels) in enumerate(test_loader):
if device != torch.device("cpu"):
images, labels = images.to(device), labels.to(device)
pred = model(images)
eval_loss.update(criterion(pred, labels))
eval_accuracy.update(accuracy(pred, labels))
print(
f"Evaluation(round {epoch}): Eval Loss:{eval_loss.avg:.4f}, "
f"Accuracy:{eval_accuracy.avg * 100:.2f}%"
)
return eval_loss.avg, eval_accuracy.avg
if __name__ == "__main__":
args = get_params().parse_args()
torch.manual_seed(args.seed)
# set num_clients = 1 to make sure there's 1 train_loader
args.num_clients = 1
device_map, train_loaders, test_loader = preprocess(args)
train_loader = train_loaders[0]
device = device_map["server"]
model, criterion, optimizer, scheduler, grad_estimator = prepare_settings(args, device)
checkpoint = CheckPoint(args, model, optimizer, grad_estimator)
args_str = get_args_str(args) + "-" + model.model_name
if args.log_to_tensorboard:
tensorboard_sub_folder = args_str + "-" + model_helpers.get_current_datetime_str()
writer = SummaryWriter(
path.join(
"tensorboards",
args.dataset,
args.log_to_tensorboard,
tensorboard_sub_folder,
)
)
sparsity_dict = use_sparsity_dict(args, model.model_name)
for epoch in range(args.epoch):
if sparsity_dict is not None and epoch % args.mask_shuffle_interval == 0:
raise NotImplementedError("We no longer support pruning mask.")
train_loss, train_accuracy = train_model(epoch)
if args.log_to_tensorboard:
writer.add_scalar("Loss/train", train_loss, epoch)
writer.add_scalar("Accuracy/train", train_accuracy, epoch)
eval_loss, eval_accuracy = eval_model(epoch)
if args.log_to_tensorboard:
writer.add_scalar("Loss/test", eval_loss, epoch)
writer.add_scalar("Accuracy/test", eval_accuracy, epoch)
if checkpoint.should_update(eval_loss, eval_accuracy, epoch):
checkpoint.save(
args_str + "-" + model_helpers.get_current_datetime_str(),
epoch,
subfolder=args.log_to_tensorboard,
)
if args.log_to_tensorboard:
writer.close()