forked from ryujaehun/pytorch-gpu-benchmark
-
Notifications
You must be signed in to change notification settings - Fork 1
/
benchmark_models.py
380 lines (344 loc) · 15.2 KB
/
benchmark_models.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
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
"""Compare speed of different models with batch size 12"""
import torch
import torchvision.models as models
import platform
import psutil
import torch.nn as nn
import datetime
import time
import os
import pandas as pd
import argparse
from torch.utils.data import Dataset, DataLoader
import json
import sys
from dataclasses import dataclass
@dataclass
class BenchmarkModelData:
model_desc: str
model_set: dict
torch.backends.cudnn.benchmark = True
# https://discuss.pytorch.org/t/what-does-torch-backends-cudnn-benchmark-do/5936
# This flag allows you to enable the inbuilt cudnn auto-tuner to find the best algorithm to use for your hardware.
# If you check it using the profile tool, the cnn method such as winograd, fft, etc. is used for the first iteration and the best operation is selected for the device.
# mnasnet0_5,mnasnet0_75
# resnet18,resnet34,resnet50
# densenet121,
# squeezenet1_0,
# vgg11,vgg11_bn,
# mobilenet_v3_small,
# shufflenet_v2_x0_5,shufflenet_v2_x1_0, shufflenet_v2_x1_5,shufflenet_v2_x2_0
MODEL_LIST_MINIMAL = {
models.mnasnet: ["mnasnet0_5", "mnasnet0_75"],
models.resnet: ["resnet18"],
models.densenet: ["densenet121"],
models.squeezenet: ["squeezenet1_0"],
models.vgg: ["vgg11"],
models.mobilenet: ["mobilenet_v3_small"],
}
PRECISION_LIST_MINIMAL = ["float"]
# mnasnet0_5,mnasnet0_75,mnasnet1_0,mnasnet1_3
# resnet18,resnet34,resnet50,resnet101,resnet152,resnext50_32x4d,resnext101_32x8d,resnext101_64x4d,wide_resnet50_2,wide_resnet101_2,
# densenet121,densenet161,densenet169,densenet201
# squeezenet1_0,squeezenet1_1,
# vgg11,vgg11_bn,vgg13,vgg13_bn,vgg16,vgg16_bn,vgg19,vgg19_bn,
# mobilenet_v3_large,mobilenet_v3_small,
# shufflenet_v2_x0_5,shufflenet_v2_x1_0, shufflenet_v2_x1_5,shufflenet_v2_x2_0
MODEL_LIST_MEDIUM = {
models.mnasnet: ["mnasnet0_5", "mnasnet0_75", "mnasnet1_0"],
models.resnet: ["resnet18", "resnet34", "resnet50", "resnet101", "resnet152"],
models.densenet: ["densenet121"],
models.squeezenet: ["squeezenet1_0", "squeezenet1_1"],
models.vgg: ["vgg11", "vgg11_bn", "vgg13", "vgg13_bn", "vgg16", "vgg16_bn"],
models.mobilenet: ["mobilenet_v3_large", "mobilenet_v3_small"],
models.shufflenetv2: ["shufflenet_v2_x0_5", "shufflenet_v2_x1_5"],
}
PRECISION_LIST_MEDIUM = ["float", "half"]
# mnasnet0_5,mnasnet0_75,mnasnet1_0,mnasnet1_3
# resnet18,resnet34,resnet50,resnet101,resnet152,resnext50_32x4d,resnext101_32x8d,resnext101_64x4d,wide_resnet50_2,wide_resnet101_2,
# densenet121,densenet161,densenet169,densenet201
# squeezenet1_0,squeezenet1_1,
# vgg11,vgg11_bn,vgg13,vgg13_bn,vgg16,vgg16_bn,vgg19,vgg19_bn,
# mobilenet_v3_large,mobilenet_v3_small,
# shufflenet_v2_x0_5,shufflenet_v2_x1_0, shufflenet_v2_x1_5,shufflenet_v2_x2_0
MODEL_LIST_FULL = {
models.mnasnet: models.mnasnet.__all__[1:],
models.resnet: models.resnet.__all__[1:],
models.densenet: models.densenet.__all__[1:],
models.squeezenet: models.squeezenet.__all__[1:],
models.vgg: models.vgg.__all__[1:],
models.mobilenet: models.mobilenet.mv3_all[1:],
models.shufflenetv2: models.shufflenetv2.__all__[1:],
}
PRECISION_LIST_FULL = ["float", "half", "double"]
# For post-voltaic architectures, there is a possibility to use tensor-core at half precision.
# Due to the gradient overflow problem, apex is recommended for practical use.
# Training settings
parser = argparse.ArgumentParser(description="PyTorch Benchmarking")
parser.add_argument(
"--WARM_UP", "-w", type=int, default=5, required=False, help="Num of warm up"
)
parser.add_argument(
"--NUM_TEST", "-n", type=int, default=50, required=False, help="Num of Test"
)
parser.add_argument(
"--BATCH_SIZE", "-b", type=int, default=12, required=False, help="Num of batch size"
)
parser.add_argument(
"--NUM_CLASSES", "-c", type=int, default=1000, required=False, help="Num of class"
)
parser.add_argument(
"--GPU_COUNT", "-g", type=int, default=1, required=False, help="Number of gpus used in test"
)
parser.add_argument(
"--GPU_INDEX", "-i", type=int, default=-1, required=False, help="Index for the used gpu"
)
parser.add_argument(
"--folder",
"-f",
type=str,
default="new_results",
required=False,
help="folder to save results",
)
class RandomDataset(Dataset):
def __init__(self, length):
self.len = length
self.data = torch.randn(3, 224, 224, length)
def __getitem__(self, index):
return self.data[:, :, :, index]
def __len__(self):
return self.len
def train(precision="single", gpu_index=-1, benchmark_model=MODEL_LIST_MINIMAL):
"""use fake image for training speed test"""
if gpu_index >= 0:
target = torch.LongTensor(args.BATCH_SIZE).random_(args.NUM_CLASSES).cuda(gpu_index)
else:
target = torch.LongTensor(args.BATCH_SIZE).random_(args.NUM_CLASSES).cuda()
criterion = nn.CrossEntropyLoss()
benchmark = {}
for model_type in benchmark_model.keys():
#print("train model_type: " + model_type)
for model_name in benchmark_model[model_type]:
print("model_name: " + model_name)
for model_type in benchmark_model.keys():
for model_name in benchmark_model[model_type]:
if model_name[-8:] == '_Weights': continue
torch_device_name = "cuda"
if (gpu_index >= 0):
torch_device_name = "cuda:" + str(gpu_index)
torch.cuda.set_device(gpu_index)
print("torch_device_name: " + torch_device_name)
model = getattr(model_type, model_name)()
if args.GPU_COUNT > 1:
model = nn.DataParallel(model, device_ids=range(args.GPU_COUNT))
model = getattr(model, precision)()
torch_device = torch.device(torch_device_name)
model = model.to(torch_device)
durations = []
print(f"Benchmarking Training {precision} precision type {model_name} ")
for step, img in enumerate(rand_loader):
img = getattr(img, precision)()
torch.cuda.synchronize()
start = time.time()
model.zero_grad()
prediction = model(img.to("cuda"))
loss = criterion(prediction, target)
loss.backward()
torch.cuda.synchronize()
end = time.time()
if step >= args.WARM_UP:
durations.append((end - start) * 1000)
print(
f"{model_name} model average train time : {sum(durations)/len(durations)}ms"
)
del model
torch.cuda.empty_cache()
benchmark[model_name] = durations
print(torch.cuda.memory_summary())
return benchmark
def inference(precision="float", gpu_index=-1, benchmark_model=MODEL_LIST_MINIMAL):
benchmark = {}
with torch.no_grad():
for model_type in benchmark_model.keys():
for model_name in benchmark_model[model_type]:
if model_name[-8:] == '_Weights': continue
torch_device_name = "cuda"
if (gpu_index >= 0):
torch_device_name = "cuda:" + str(gpu_index)
torch.cuda.set_device(gpu_index)
print("torch_device_name: " + torch_device_name)
torch_device = torch.device(torch_device_name)
model = getattr(model_type, model_name)()
if args.GPU_COUNT > 1:
model = nn.DataParallel(model, device_ids=range(args.GPU_COUNT))
model = getattr(model, precision)()
model = model.to(torch_device)
model.eval()
durations = []
print(
f"Benchmarking Inference {precision} precision type {model_name} "
)
for step, img in enumerate(rand_loader):
img = getattr(img, precision)()
torch.cuda.synchronize()
start = time.time()
model(img.to(torch_device))
torch.cuda.synchronize()
end = time.time()
if step >= args.WARM_UP:
durations.append((end - start) * 1000)
print(
f"{model_name} model average inference time : {sum(durations)/len(durations)}ms"
)
del model
torch.cuda.empty_cache()
benchmark[model_name] = durations
print(torch.cuda.memory_summary())
return benchmark
f"{platform.uname()}\n{psutil.cpu_freq()}\ncpu_count: {psutil.cpu_count()}\nmemory_available: {psutil.virtual_memory().available}"
if __name__ == "__main__":
args = parser.parse_args()
args.BATCH_SIZE *= args.GPU_COUNT
gpu_count = args.GPU_COUNT
gpu_index = args.GPU_INDEX
print("gpu_index: " + str(gpu_index))
print("gpu_count: " + str(gpu_count))
print("BATCH_SIZE: " + str(args.BATCH_SIZE))
rand_loader = DataLoader(
dataset=RandomDataset(args.BATCH_SIZE * (args.WARM_UP + args.NUM_TEST)),
batch_size=args.BATCH_SIZE,
shuffle=False,
num_workers=8,
)
# get available memory
gpu_mem_total = 0
gpu_mem_used = 0
if (gpu_index >= 0):
device_name = str(torch.cuda.get_device_name(gpu_index))
mem_tuple = torch.cuda.mem_get_info(gpu_index)
gpu_mem_total = mem_tuple[0]
gpu_mem_free = mem_tuple[1]
else:
device_name = str(torch.cuda.get_device_name(0))
# search which gpu has smallest amount of memory
for ii in range (gpu_count):
mem_tuple = torch.cuda.mem_get_info(ii)
if (ii == 0):
gpu_mem_total = mem_tuple[0]
gpu_mem_free = mem_tuple[1]
else:
if (mem_tuple[0] < gpu_mem_total):
gpu_mem_total = mem_tuple[0]
gpu_mem_free = mem_tuple[1]
# convert memory sizes to gigabytes (gb)
gpu_mem_total = gpu_mem_total / (1048576.0 * 1024)
gpu_mem_free = gpu_mem_free / (1048576.0 * 1024)
gpu_mem_used = gpu_mem_total - gpu_mem_free
# select which set of benchmarks and precisions to run
# depending from the gpu memory available. (to avoid out of memory errors)
gpu_benchmark_models_name = "MINIMAL"
modeldata_float = BenchmarkModelData("MEDIUM", MODEL_LIST_MEDIUM)
modeldata_half = BenchmarkModelData("MINIMAL", MODEL_LIST_MINIMAL)
modeldata_double = BenchmarkModelData("MINIMAL", MODEL_LIST_MINIMAL)
# run medium list also with integrated graphic cards
precisions = PRECISION_LIST_MEDIUM
benchmark_model_dict = {}
# if gpu is AMD's integrated graphic card, run only the minime set of benchmarks
if device_name == "AMD Radeon Graphics":
precisions = PRECISION_LIST_MINIMAL
benchmark_model_dict["float"] = BenchmarkModelData("MINIMAL", MODEL_LIST_MINIMAL)
#benchmark_model_dict["half"] = BenchmarkModelData("MINIMAL", MODEL_LIST_MINIMAL)
gpu_benchmark_models_name = "MINIMAL"
else:
if (gpu_mem_free <= 8):
precisions = PRECISION_LIST_MEDIUM
benchmark_model_dict["float"] = BenchmarkModelData("MINIMAL", MODEL_LIST_MINIMAL)
benchmark_model_dict["half"] = BenchmarkModelData("MINIMAL", MODEL_LIST_MINIMAL)
#benchmark_model_dict["double"] = BenchmarkModelData("MINIMAL", MODEL_LIST_MINIMAL)
gpu_benchmark_models_name = "MINIMAL"
elif (gpu_mem_free > 8) and (gpu_mem_free <= 10):
precisions = PRECISION_LIST_FULL
benchmark_model_dict["float"] = BenchmarkModelData("MEDIUM", MODEL_LIST_MEDIUM)
benchmark_model_dict["half"] = BenchmarkModelData("MEDIUM", MODEL_LIST_MEDIUM)
benchmark_model_dict["double"] = BenchmarkModelData("MEDIUM", MODEL_LIST_MEDIUM)
gpu_benchmark_models_name = "MEDIUM"
else:
precisions = PRECISION_LIST_FULL
benchmark_model_dict["float"] = BenchmarkModelData("FULL", MODEL_LIST_FULL)
benchmark_model_dict["half"] = BenchmarkModelData("FULL", MODEL_LIST_FULL)
benchmark_model_dict["double"] = BenchmarkModelData("FULL", MODEL_LIST_FULL)
gpu_benchmark_models_name = "FULL"
device_name = f"{device_name}"
if (args.GPU_COUNT > 1):
device_name = device_name + str(gpu_count) + "X"
device_name = device_name.replace(" ", "_")
device_file_name = device_name + "_"
print("device_name: " + device_name)
print("mem free: " + str(gpu_mem_free))
if (gpu_index >= 0):
folder_name = args.folder + "/" + str(gpu_index) + "/" + device_name
else:
folder_name = args.folder + "/" + str(gpu_count) + "X"
print("folder_name: " + folder_name)
system_configs = f"{platform.uname()}\n\
{psutil.cpu_freq()}\n\
cpu_count: {psutil.cpu_count()}\n\
memory_available: {psutil.virtual_memory().available}\n\
gpu_benchmark_models_description: {gpu_benchmark_models_name}"
gpu_configs = [
gpu_count,
torch.__version__,
torch.version.hip,
torch.version.cuda,
torch.backends.cudnn.version(),
device_name,
gpu_mem_total,
gpu_mem_used,
gpu_benchmark_models_name,
]
gpu_configs = list(map(str, gpu_configs))
temp = [
"GPU_Count: ",
"Torch_Version : ",
"ROCM_Version: ",
"CUDA_Version: ",
"Cudnn_Version: ",
"Device_Name: ",
"GPU_Mem_Total_GB: ",
"GPU_Mem_Free_GB: ",
"Benchmark_Model_Size: ",
]
os.makedirs(folder_name, exist_ok=True)
with open(os.path.join(folder_name, "config.json"), "w") as f:
json.dump(vars(args), f, indent=2)
now = datetime.datetime.now()
start_time = now.strftime("%Y/%m/%d %H:%M:%S")
print(f"benchmark start : {start_time}")
for idx, value in enumerate(zip(temp, gpu_configs)):
gpu_configs[idx] = "".join(value)
print(gpu_configs[idx])
print(system_configs)
with open(os.path.join(folder_name, "system_info.txt"), "w") as f:
f.writelines(f"benchmark start : {start_time}\n")
f.writelines("system_configs\n\n")
f.writelines(system_configs)
f.writelines("\ngpu_configs\n\n")
f.writelines(s + "\n" for s in gpu_configs)
for precision in precisions:
benchmark_model_data = benchmark_model_dict[precision]
print("precision: " + precision + ", set: " + benchmark_model_data.model_desc)
train_result = train(precision, gpu_index, benchmark_model_data.model_set)
train_result_df = pd.DataFrame(train_result)
path = f"{folder_name}/{device_file_name}_{precision}_model_train_benchmark.csv"
train_result_df.to_csv(path, index=False)
inference_result = inference(precision, gpu_index, benchmark_model_data.model_set)
inference_result_df = pd.DataFrame(inference_result)
path = f"{folder_name}/{device_file_name}_{precision}_model_inference_benchmark.csv"
inference_result_df.to_csv(path, index=False)
# finish the benchmarks
now = datetime.datetime.now()
end_time = now.strftime("%Y/%m/%d %H:%M:%S")
print(f"benchmark end : {end_time}")
with open(os.path.join(folder_name, "system_info.txt"), "a") as f:
f.writelines(f"benchmark end : {end_time}\n")