forked from SonyResearch/MocoSFL
-
Notifications
You must be signed in to change notification settings - Fork 0
/
overhead_calculator.py
167 lines (137 loc) · 5.92 KB
/
overhead_calculator.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
# %%
from pyexpat import model
import torch
import torch.nn as nn
import numpy as np
from models import resnet
from models import vgg
import subprocess
import sys
try:
from fvcore.nn import FlopCountAnalysis
except ImportError:
subprocess.check_call([sys.executable, "-m", "pip", "install", 'fvcore'])
finally:
from fvcore.nn import FlopCountAnalysis
from fvcore.nn import flop_count_table
#archs
arch="ResNet18"
#Dataset related
input_channel_size = 3
data_size=32
num_class=10
#Scheme related # Change Here
scheme="SFL" #SFL/FL
TAResSFL_enable = False
cutlayer=5
adds_bottleneck = False
bottleneck_option="C4S2"
#Moco-related inputs
if scheme=="SFL":
moco_version="V2"
K_dim=1024
sync_frequency = 1
num_epoch_per_client=200
num_data_per_client=50
batch_size=1
else:
moco_version="largeV2"
K_dim=2048
sync_frequency = 5 # how many client epoch per sync
num_epoch_per_client=500
num_data_per_client=10000
batch_size=128
# server-side computations
num_clients = 1000
client_sampling_ratio = 0.1
if scheme=="FL":
adds_bottleneck = False
if "ResNet" in arch or "resnet" in arch:
if "resnet" in arch:
arch = "ResNet" + arch.split("resnet")[-1]
create_arch = getattr(resnet, arch)
elif "vgg" in arch:
create_arch = getattr(vgg, arch)
#get model - use a larger classifier, as in Zhuang et al. Divergence-aware paper
global_model = create_arch(cutting_layer=cutlayer, num_client = 1, num_class=K_dim, group_norm=True, input_size= data_size,
adds_bottleneck=adds_bottleneck, bottleneck_option=bottleneck_option)
if moco_version == "largeV2": # This one uses a larger classifier, same as in Zhuang et al. Divergence-aware paper
classifier_list = [nn.Linear(512 * global_model.expansion, 4096),
nn.BatchNorm1d(4096),
nn.ReLU(True),
nn.Linear(4096, K_dim)]
global_model.classifier = nn.Sequential(*classifier_list)
elif "V2" in moco_version:
classifier_list = [nn.Linear(512 * global_model.expansion, K_dim * global_model.expansion),
nn.ReLU(True),
nn.Linear(K_dim * global_model.expansion, K_dim)]
global_model.classifier = nn.Sequential(*classifier_list)
global_model.merge_classifier_cloud()
if scheme=="FL":
if global_model.get_num_of_cloud_layer() != 0:
global_model.resplit(0)
if scheme == "SFL":
latent_vector_total_size=np.prod(global_model.get_smashed_data_size(1, data_size))
weight_param_size = 0
for key in global_model.local_list[0].state_dict().keys():
weight_param_size += np.prod(global_model.local_list[0].state_dict()[key].size())
communication_overhead_weight = num_epoch_per_client//sync_frequency * weight_param_size
if scheme == "FL":
communication_overhead_weight_latent_vector = 0
elif scheme == "SFL":
communication_overhead_weight_latent_vector = 2 * num_epoch_per_client * num_data_per_client * latent_vector_total_size
if scheme == "SFL" and TAResSFL_enable:
communication_overhead_weight = 0.0
communication_overhead_weight_latent_vector = communication_overhead_weight_latent_vector/2
communication_overhead = communication_overhead_weight + communication_overhead_weight_latent_vector
print("===============================")
print(f"Model weight communication overhead: {communication_overhead_weight*4/1024/1024:.2f} MB")
print(f"Latent vector communication overhead: {communication_overhead_weight_latent_vector*4/1024/1024:.2f} MB")
print(f"Total communication overhead: {communication_overhead*4/1024/1024:.2f} MB")
print("===============================")
exit(0)
#get_memory_usage
global_model.local_list[0].cuda()
global_model.cloud.cuda()
noise_input = torch.ones([batch_size, input_channel_size, data_size, data_size])
noise_label = torch.ones(global_model.get_smashed_data_size(batch_size, data_size))
criterion = nn.MSELoss()
noise_input = noise_input.cuda()
noise_label = noise_label.cuda()
params = list(global_model.local_list[0].parameters())
optimizer = torch.optim.SGD(params, lr=0.02, momentum=0.9, weight_decay=5e-4)
if scheme == "SFL" and TAResSFL_enable:
with torch.no_grad():
output = global_model.local_list[0](noise_input)
print("Total CUDA Memory Allocated for inference: %.2f MB"%(torch.cuda.memory_allocated(0)/1024/1024))
#GPU warmup
for i in range(5):
optimizer.zero_grad()
output = global_model.local_list[0](noise_input)
f_loss = criterion(output, noise_label)
if i == 4:
print("Total CUDA Memory Allocated for training: %.2f MB"%(torch.cuda.memory_allocated(0)/1024/1024))
f_loss.backward()
optimizer.step()
noise_input = torch.ones([1, input_channel_size, data_size, data_size])
noise_input = noise_input.cuda()
print("===============================")
flops = FlopCountAnalysis(global_model.local_list[0], noise_input)
if scheme == "SFL":
noise_input = torch.ones(global_model.get_smashed_data_size(1, data_size))
noise_input = noise_input.cuda()
print("===============================")
server_flops = 2 * FlopCountAnalysis(global_model.cloud, noise_input).total() + weight_param_size
else:
server_flops = weight_param_size
if scheme == "SFL" and TAResSFL_enable: # if TAResSFL_enable, then no training, no momentum, if not, then + backward + momemtum forward.
print(f"FLOPs/image: {flops.total()/1024/1024:.2f} M")
print(f"Total FLOPs: {num_epoch_per_client * num_data_per_client * flops.total()/1024/1024/1024:.2f} G")
else:
print(f"FLOPs/image: {3* flops.total()/1024/1024:.2f} M")
print(f"Total FLOPs: {3*num_epoch_per_client * num_data_per_client * flops.total()/1024/1024/1024:.2f} G")
print(f"Total server-side FLOPs: {num_clients * num_epoch_per_client // sync_frequency * server_flops/1024/1024/1024:.2f} G")
# print(f"Total FLOPs (by operator): {flops.by_operator()} M")
# print(f"Total FLOPs (by module): {flops.by_module()} M")
# print(flop_count_table(flops))
print("===============================")