-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathMNIST_train_test_grad_acc.py
247 lines (190 loc) · 12.4 KB
/
MNIST_train_test_grad_acc.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
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
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision
from torch.utils.data import DataLoader
import torchvision.transforms as transforms
import numpy as np
import time
from torchsummary import summary
from utils import clear_gpu_mem_util
from utils import check_gpu_info_queue
from utils import trig_GPU_read
from utils import get_info_from_GPU_queue
from models.models import get_model
from models.torchSummaryWrapper import get_torchSummaryWrapper
def train_and_test( data_path,
input_size,
learning_rate,
n_epochs,
num_workers,
bn_or_gn,
en_grad_checkpointing,
training_params,
optimizer_types,
event_start_read_GPU_info,
queue_gpu_info,
queue_training_results, ):
tot_iter = len( training_params ) * len(optimizer_types)
train_acc_vec = np.zeros( ( tot_iter, n_epochs ) )
test_acc_vec = np.zeros( ( tot_iter, n_epochs ) )
train_time_vec = np.zeros( ( tot_iter, n_epochs ) )
test_time_vec = np.zeros( ( tot_iter, n_epochs ) )
train_gpu_mem_usage = np.zeros( ( tot_iter, n_epochs ) )
test_gpu_mem_usage = np.zeros( ( tot_iter, n_epochs ) )
train_gpu_util = np.zeros( ( tot_iter, n_epochs ) )
test_gpu_util = np.zeros( ( tot_iter, n_epochs ) )
iter_index = 0
for training_param in training_params:
input_expand_ratio = training_param[0]
batch_size = training_param[1]
optimizing_batch = training_param[2]
transform = transforms.Compose([
transforms.Resize( [input_size*input_expand_ratio, input_size*input_expand_ratio ] ),
transforms.ToTensor(), # Convert PIL images to tensors
transforms.Normalize((0.5,), (0.5,)) # Normalize the image data to the range [-1, 1]
])
train_dataset = torchvision.datasets.MNIST( root = data_path,
train = True,
download = True,
transform = transform )
test_dataset = torchvision.datasets.MNIST( root = data_path,
train = False,
download = True,
transform = transform )
for optimizer_type in optimizer_types:
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
if( device == 'gpu' ):
torch.cuda.empty_cache()
# device = 'cpu'
model = get_model( input_expand_ratio, bn_or_gn, en_grad_checkpointing ).to(device)
criterion = nn.CrossEntropyLoss()
if(en_grad_checkpointing==False):
summary(model, (1, input_size*input_expand_ratio, input_size*input_expand_ratio) )
else:
summary(get_torchSummaryWrapper( model ), (1, input_size*input_expand_ratio, input_size*input_expand_ratio) )
if(optimizer_type == 'Adam'):
optimizer = optim.Adam(model.parameters(), lr=learning_rate)
else:
optimizer = optim.SGD(model.parameters(), lr=learning_rate)
# Training loop
print("-" * 80)
print("Started Training and Test " + str(iter_index) + '/' + str(tot_iter-1) + " for batch size " + str(batch_size) +
", optimizing batch " + str(optimizing_batch) + ", optimizer type " + optimizer_type )
for epoch in range(n_epochs):
correct_train = 0
total_train = 0
correct_test = 0
total_test = 0
gpu_mem_usage_max_train = 0
gpu_util_max_train = 0
gpu_mem_usage_max_test = 0
gpu_util_max_test = 0
device_name = []
device_mem_cap = 0
start_time_train = time.perf_counter()
train_dataloader = DataLoader( dataset = train_dataset,
batch_size = batch_size,
shuffle = True,
pin_memory = True,
num_workers = num_workers )
for i, (images, labels) in enumerate(train_dataloader):
trig_GPU_read( queue_gpu_info, event_start_read_GPU_info )
if(en_grad_checkpointing==True):
images.requires_grad = True
images, labels = images.to(device), labels.to(device)
# Forward pass
outputs = model(images)
loss = criterion(outputs, labels)
loss = loss / optimizing_batch
# Backward and optimize
loss.backward()
_, predicted = torch.max(outputs.data, 1)
total_train += labels.size(0)
correct_train += (predicted == labels).sum().cpu().detach().numpy()
acc_train = correct_train / total_train
if( (i+1) % optimizing_batch == 0 or (i+1) == len(train_dataloader) ):
optimizer.step()
optimizer.zero_grad()
data_from_GPU_queue = get_info_from_GPU_queue( queue_gpu_info, event_start_read_GPU_info)
if(data_from_GPU_queue != None):
device_name = data_from_GPU_queue[0]
device_mem_cap = data_from_GPU_queue[1]
gpu_mem_usage_max_train = max(gpu_mem_usage_max_train, data_from_GPU_queue[2])
gpu_util_max_train = max(gpu_util_max_train, data_from_GPU_queue[3])
if( ( (i*batch_size) % 10000 ) > ( ((i+1)*batch_size) % 10000 ) or (i+1) == len(train_dataloader) ):
print("Train Epoch {}/{} Batch {}/{} LR {:.6f} Loss {:.6f} CorPred {}/{} Acc {:.6f} {} Mem Used {}/{} GPU Util {}"
.format( epoch,
n_epochs-1,
i,
len(train_dataloader)-1,
learning_rate,
loss.cpu().detach().numpy(),
correct_train,
total_train,
acc_train,
device_name,
gpu_mem_usage_max_train,
device_mem_cap,
gpu_util_max_train, ) )
check_gpu_info_queue(queue_gpu_info)
train_acc_vec[iter_index, epoch] = acc_train
end_time_train = time.perf_counter()
train_time_vec[iter_index, epoch] = end_time_train - start_time_train
train_gpu_mem_usage[iter_index, epoch] = gpu_mem_usage_max_train
train_gpu_util[iter_index, epoch] = gpu_util_max_train
start_time_test = time.perf_counter()
test_dataloader = DataLoader( dataset = test_dataset,
batch_size = batch_size,
shuffle = False,
pin_memory = True,
num_workers = num_workers )
model.eval() # Sets the model to evaluation mode
with torch.no_grad():
for i, (images, labels) in enumerate(test_dataloader):
trig_GPU_read( queue_gpu_info, event_start_read_GPU_info )
images, labels = images.to(device), labels.to(device)
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
total_test += labels.size(0)
correct_test += (predicted == labels).sum().cpu().detach().numpy()
acc_test = correct_test / total_test
data_from_GPU_queue = get_info_from_GPU_queue( queue_gpu_info, event_start_read_GPU_info)
if(data_from_GPU_queue != None):
device_name = data_from_GPU_queue[0]
device_mem_cap = data_from_GPU_queue[1]
gpu_mem_usage_max_test = max(gpu_mem_usage_max_train, data_from_GPU_queue[2])
gpu_util_max_test = max(gpu_util_max_train, data_from_GPU_queue[3])
if( ( (i*batch_size) % 10000 ) > ( ((i+1)*batch_size) % 10000 ) or (i+1) == len(test_dataloader) ):
print("Test Epoch {}/{} Batch {}/{} LR {:.6f} CorPred {}/{} Acc {:.6f} {} Mem Used {}/{} GPU Util {}"
.format( epoch,
n_epochs-1,
i,
len(test_dataloader)-1,
learning_rate,
correct_test,
total_test,
acc_test,
device_name,
gpu_mem_usage_max_test,
device_mem_cap,
gpu_util_max_test, ) )
check_gpu_info_queue(queue_gpu_info)
test_acc_vec[iter_index, epoch] = acc_test
end_time_test = time.perf_counter()
test_time_vec[iter_index, epoch] = end_time_test - start_time_test
test_gpu_mem_usage[iter_index, epoch] = gpu_mem_usage_max_test
test_gpu_util[iter_index, epoch] = gpu_util_max_test
print("-" * 40)
iter_index = iter_index + 1
clear_gpu_mem_util(model, images, labels)
queue_training_results.put( ( train_acc_vec,
test_acc_vec,
train_time_vec,
test_time_vec,
train_gpu_mem_usage,
test_gpu_mem_usage,
train_gpu_util,
test_gpu_util,
device_name,
device_mem_cap, ) )