Note
Click here to download the full example code
PyTorch 模型分析¶
Author: Suraj Subramanian
PyTorch includes a profiler API that is useful to identify the time and memory costs of various PyTorch operations in your code. Profiler can be easily integrated in your code, and the results can be printed as a table or returned in a JSON trace file.
Note
Profiler supports multithreaded models. Profiler runs in the same thread as the operation but it will also profile child operators that might run in another thread. Concurrently-running profilers will be scoped to their own thread to prevent mixing of results.
Note
PyTorch 1.8 introduces the new API that will replace the older profiler API in the future releases. Check the new API at this page.
Head on over to this recipe for a quicker walkthrough of Profiler API usage.
import torch
import numpy as np
from torch import nn
import torch.autograd.profiler as profiler
Performance debugging using Profiler¶
Profiler can be useful to identify performance bottlenecks in your models. In this example, we build a custom module that performs two sub-tasks:
a linear transformation on the input, and
use the transformation result to get indices on a mask tensor.
We wrap the code for each sub-task in separate labelled context managers using
profiler.record_function("label"). In the profiler output, the
aggregate performance metrics of all operations in the sub-task will
show up under its corresponding label.
Note that using Profiler incurs some overhead, and is best used only for investigating code. Remember to remove it if you are benchmarking runtimes.
class MyModule(nn.Module):
def __init__(self, in_features: int, out_features: int, bias: bool = True):
super(MyModule, self).__init__()
self.linear = nn.Linear(in_features, out_features, bias)
def forward(self, input, mask):
with profiler.record_function("LINEAR PASS"):
out = self.linear(input)
with profiler.record_function("MASK INDICES"):
threshold = out.sum(axis=1).mean().item()
hi_idx = np.argwhere(mask.cpu().numpy() > threshold)
hi_idx = torch.from_numpy(hi_idx).cuda()
return out, hi_idx
Profile the forward pass¶
We initialize random input and mask tensors, and the model.
Before we run the profiler, we warm-up CUDA to ensure accurate
performance benchmarking. We wrap the forward pass of our module in the
profiler.profile context manager. The with_stack=True parameter appends the
file and line number of the operation in the trace.
Warning
with_stack=True incurs an additional overhead, and is better suited for investigating code.
Remember to remove it if you are benchmarking performance.
model = MyModule(500, 10).cuda()
input = torch.rand(128, 500).cuda()
mask = torch.rand((500, 500, 500), dtype=torch.double).cuda()
# warm-up
model(input, mask)
with profiler.profile(with_stack=True, profile_memory=True) as prof:
out, idx = model(input, mask)
Print profiler results¶
Finally, we print the profiler results. profiler.key_averages
aggregates the results by operator name, and optionally by input
shapes and/or stack trace events.
Grouping by input shapes is useful to identify which tensor shapes
are utilized by the model.
Here, we use group_by_stack_n=5 which aggregates runtimes by the
operation and its traceback (truncated to the most recent 5 events), and
display the events in the order they are registered. The table can also
be sorted by passing a sort_by argument (refer to the
docs for
valid sorting keys).
Note
When running profiler in a notebook, you might see entries like <ipython-input-18-193a910735e8>(13): forward
instead of filenames in the stacktrace. These correspond to <notebook-cell>(line number): calling-function.
print(prof.key_averages(group_by_stack_n=5).table(sort_by='self_cpu_time_total', row_limit=5))
"""
(Some columns are omitted)
------------- ------------ ------------ ------------ ---------------------------------
Name Self CPU % Self CPU Self CPU Mem Source Location
------------- ------------ ------------ ------------ ---------------------------------
MASK INDICES 87.88% 5.212s -953.67 Mb /mnt/xarfuse/.../torch/au
<ipython-input-...>(10): forward
/mnt/xarfuse/.../torch/nn
<ipython-input-...>(9): <module>
/mnt/xarfuse/.../IPython/
aten::copy_ 12.07% 715.848ms 0 b <ipython-input-...>(12): forward
/mnt/xarfuse/.../torch/nn
<ipython-input-...>(9): <module>
/mnt/xarfuse/.../IPython/
/mnt/xarfuse/.../IPython/
LINEAR PASS 0.01% 350.151us -20 b /mnt/xarfuse/.../torch/au
<ipython-input-...>(7): forward
/mnt/xarfuse/.../torch/nn
<ipython-input-...>(9): <module>
/mnt/xarfuse/.../IPython/
aten::addmm 0.00% 293.342us 0 b /mnt/xarfuse/.../torch/nn
/mnt/xarfuse/.../torch/nn
/mnt/xarfuse/.../torch/nn
<ipython-input-...>(8): forward
/mnt/xarfuse/.../torch/nn
aten::mean 0.00% 235.095us 0 b <ipython-input-...>(11): forward
/mnt/xarfuse/.../torch/nn
<ipython-input-...>(9): <module>
/mnt/xarfuse/.../IPython/
/mnt/xarfuse/.../IPython/
----------------------------- ------------ ---------- ----------------------------------
Self CPU time total: 5.931s
"""
Improve memory performance¶
Note that the most expensive operations - in terms of memory and time -
are at forward (10) representing the operations within MASK INDICES. Let’s try to
tackle the memory consumption first. We can see that the .to()
operation at line 12 consumes 953.67 Mb. This operation copies mask to the CPU.
mask is initialized with a torch.double datatype. Can we reduce the memory footprint by casting
it to torch.float instead?
model = MyModule(500, 10).cuda()
input = torch.rand(128, 500).cuda()
mask = torch.rand((500, 500, 500), dtype=torch.float).cuda()
# warm-up
model(input, mask)
with profiler.profile(with_stack=True, profile_memory=True) as prof:
out, idx = model(input, mask)
print(prof.key_averages(group_by_stack_n=5).table(sort_by='self_cpu_time_total', row_limit=5))
"""
(Some columns are omitted)
----------------- ------------ ------------ ------------ --------------------------------
Name Self CPU % Self CPU Self CPU Mem Source Location
----------------- ------------ ------------ ------------ --------------------------------
MASK INDICES 93.61% 5.006s -476.84 Mb /mnt/xarfuse/.../torch/au
<ipython-input-...>(10): forward
/mnt/xarfuse/ /torch/nn
<ipython-input-...>(9): <module>
/mnt/xarfuse/.../IPython/
aten::copy_ 6.34% 338.759ms 0 b <ipython-input-...>(12): forward
/mnt/xarfuse/.../torch/nn
<ipython-input-...>(9): <module>
/mnt/xarfuse/.../IPython/
/mnt/xarfuse/.../IPython/
aten::as_strided 0.01% 281.808us 0 b <ipython-input-...>(11): forward
/mnt/xarfuse/.../torch/nn
<ipython-input-...>(9): <module>
/mnt/xarfuse/.../IPython/
/mnt/xarfuse/.../IPython/
aten::addmm 0.01% 275.721us 0 b /mnt/xarfuse/.../torch/nn
/mnt/xarfuse/.../torch/nn
/mnt/xarfuse/.../torch/nn
<ipython-input-...>(8): forward
/mnt/xarfuse/.../torch/nn
aten::_local 0.01% 268.650us 0 b <ipython-input-...>(11): forward
_scalar_dense /mnt/xarfuse/.../torch/nn
<ipython-input-...>(9): <module>
/mnt/xarfuse/.../IPython/
/mnt/xarfuse/.../IPython/
----------------- ------------ ------------ ------------ --------------------------------
Self CPU time total: 5.347s
"""
The CPU memory footprint for this operation has halved.
Improve time performance¶
While the time consumed has also reduced a bit, it’s still too high.
Turns out copying a matrix from CUDA to CPU is pretty expensive!
The aten::copy_ operator in forward (12) copies mask to CPU
so that it can use the NumPy argwhere function. aten::copy_ at forward(13)
copies the array back to CUDA as a tensor. We could eliminate both of these if we use a
torch function nonzero() here instead.
class MyModule(nn.Module):
def __init__(self, in_features: int, out_features: int, bias: bool = True):
super(MyModule, self).__init__()
self.linear = nn.Linear(in_features, out_features, bias)
def forward(self, input, mask):
with profiler.record_function("LINEAR PASS"):
out = self.linear(input)
with profiler.record_function("MASK INDICES"):
threshold = out.sum(axis=1).mean()
hi_idx = (mask > threshold).nonzero(as_tuple=True)
return out, hi_idx
model = MyModule(500, 10).cuda()
input = torch.rand(128, 500).cuda()
mask = torch.rand((500, 500, 500), dtype=torch.float).cuda()
# warm-up
model(input, mask)
with profiler.profile(with_stack=True, profile_memory=True) as prof:
out, idx = model(input, mask)
print(prof.key_averages(group_by_stack_n=5).table(sort_by='self_cpu_time_total', row_limit=5))
"""
(Some columns are omitted)
-------------- ------------ ------------ ------------ ---------------------------------
Name Self CPU % Self CPU Self CPU Mem Source Location
-------------- ------------ ------------ ------------ ---------------------------------
aten::gt 57.17% 129.089ms 0 b <ipython-input-...>(12): forward
/mnt/xarfuse/.../torch/nn
<ipython-input-...>(25): <module>
/mnt/xarfuse/.../IPython/
/mnt/xarfuse/.../IPython/
aten::nonzero 37.38% 84.402ms 0 b <ipython-input-...>(12): forward
/mnt/xarfuse/.../torch/nn
<ipython-input-...>(25): <module>
/mnt/xarfuse/.../IPython/
/mnt/xarfuse/.../IPython/
INDEX SCORE 3.32% 7.491ms -119.21 Mb /mnt/xarfuse/.../torch/au
<ipython-input-...>(10): forward
/mnt/xarfuse/.../torch/nn
<ipython-input-...>(25): <module>
/mnt/xarfuse/.../IPython/
aten::as_strided 0.20% 441.587us 0 b <ipython-input-...>(12): forward
/mnt/xarfuse/.../torch/nn
<ipython-input-...>(25): <module>
/mnt/xarfuse/.../IPython/
/mnt/xarfuse/.../IPython/
aten::nonzero
_numpy 0.18% 395.602us 0 b <ipython-input-...>(12): forward
/mnt/xarfuse/.../torch/nn
<ipython-input-...>(25): <module>
/mnt/xarfuse/.../IPython/
/mnt/xarfuse/.../IPython/
-------------- ------------ ------------ ------------ ---------------------------------
Self CPU time total: 225.801ms
"""
Further Reading¶
We have seen how Profiler can be used to investigate time and memory bottlenecks in PyTorch models. Read more about Profiler here:
Total running time of the script: ( 0 minutes 0.000 seconds)
