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
|
import torchvision.models as models
import torch
import torch.autograd.profiler as profiler
for with_cuda in [False, True]:
model = models.resnet18()
inputs = torch.randn(5, 3, 224, 224)
sort_key = "self_cpu_memory_usage"
if with_cuda and torch.cuda.is_available():
model = model.cuda()
inputs = inputs.cuda()
sort_key = "self_cuda_memory_usage"
print("Profiling CUDA Resnet model")
else:
print("Profiling CPU Resnet model")
with profiler.profile(profile_memory=True, record_shapes=True) as prof:
with profiler.record_function("root"):
model(inputs)
print(
prof.key_averages(group_by_input_shape=True).table(
sort_by=sort_key, row_limit=-1
)
)
|
