Skip to content

[proto] Small improvement for tensor equalize op#6738

Merged
vfdev-5 merged 5 commits intopytorch:mainfrom
vfdev-5:proto-small-optim-equalize
Oct 11, 2022
Merged

[proto] Small improvement for tensor equalize op#6738
vfdev-5 merged 5 commits intopytorch:mainfrom
vfdev-5:proto-small-optim-equalize

Conversation

@vfdev-5
Copy link
Copy Markdown
Contributor

@vfdev-5 vfdev-5 commented Oct 11, 2022 •
edited
Loading 

Time benchmark: RandomEqualize (1.0,) None
V2: RandomEqualize(p=1.0) torchvision.prototype.transforms._color
Stable: RandomEqualize(p=1.0) torchvision.transforms.transforms

Main:

[- Classification transforms measurements -]
                         |  stable  |    v2
1 threads: ---------------------------------
      Tensor Image data  |  2.875   |  3.184

Times are in milliseconds (ms).

This PR:

[- Classification transforms measurements -]
                         |  stable  |    v2
1 threads: ---------------------------------
      Tensor Image data  |  2.883   |  2.874

Times are in milliseconds (ms).

Here is cprof logs to see number of calls reduction:

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
-     6000    3.918    0.001    6.354    0.001 /vision/torchvision/transforms/functional_tensor.py:870(_scale_channel)
-     6000    1.216    0.000    1.216    0.000 {built-in method torch.bincount}
-    12000    0.957    0.000    0.957    0.000 {method 'to' of 'torch._C._TensorBase' objects}
-     4000    0.360    0.000    0.360    0.000 {built-in method torch.stack}
+     6000    3.433    0.001    5.380    0.001 /vision/torchvision/prototype/transforms/functional/_color.py:186(_scale_channel)
+     6000    1.229    0.000    1.229    0.000 {built-in method torch.bincount}
+    12000    0.454    0.000    0.454    0.000 {method 'to' of 'torch._C._TensorBase' objects}
+     2000    0.187    0.000    0.187    0.000 {built-in method torch.stack}

Main (12adc54):

   660002 function calls in 7.247 seconds

   Ordered by: internal time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
     6000    3.918    0.001    6.354    0.001 /vision/torchvision/transforms/functional_tensor.py:870(_scale_channel)
     6000    1.216    0.000    1.216    0.000 {built-in method torch.bincount}
    12000    0.957    0.000    0.957    0.000 {method 'to' of 'torch._C._TensorBase' objects}
     4000    0.360    0.000    0.360    0.000 {built-in method torch.stack}
    18000    0.080    0.000    0.080    0.000 {built-in method torch.div}
     2000    0.069    0.000    6.424    0.003 /vision/torchvision/transforms/functional_tensor.py:892(<listcomp>)
     6000    0.061    0.000    0.061    0.000 {built-in method torch.nn.functional.pad}
    10000    0.043    0.000    0.043    0.000 {method 'view' of 'torch._C._TensorBase' objects}
     2000    0.040    0.000    6.984    0.003 /vision/torchvision/prototype/transforms/_transform.py:66(forward)
     2000    0.038    0.000    0.141    0.000 /usr/lib/python3.8/traceback.py:321(extract)
    10000    0.037    0.000    0.037    0.000 {built-in method posix.stat}
     6000    0.036    0.000    0.036    0.000 {method 'clamp' of 'torch._C._TensorBase' objects}
     6000    0.034    0.000    0.034    0.000 {method 'sum' of 'torch._C._TensorBase' objects}
     6000    0.031    0.000    0.031    0.000 {built-in method torch.cumsum}
     2000    0.030    0.000    0.053    0.000 /usr/lib/python3.8/traceback.py:388(format)
     2000    0.029    0.000    6.642    0.003 /vision/torchvision/transforms/functional_tensor.py:891(_equalize_single_image)
     2000    0.023    0.000    0.023    0.000 {built-in method torch.rand}
     2000    0.017    0.000    6.880    0.003 /vision/torchvision/prototype/transforms/functional/_color.py:186(equalize_image_tensor)
    54000    0.014    0.000    0.033    0.000 /usr/lib/python3.8/traceback.py:285(line)
    36000    0.012    0.000    0.012    0.000 {method 'format' of 'str' objects}

This PR:

   652002 function calls in 6.011 seconds

   Ordered by: internal time

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
     6000    3.433    0.001    5.380    0.001 /vision/torchvision/prototype/transforms/functional/_color.py:186(_scale_channel)
     6000    1.229    0.000    1.229    0.000 {built-in method torch.bincount}
    12000    0.454    0.000    0.454    0.000 {method 'to' of 'torch._C._TensorBase' objects}
     2000    0.187    0.000    0.187    0.000 {built-in method torch.stack}
    18000    0.085    0.000    0.085    0.000 {built-in method torch.div}
     6000    0.058    0.000    0.058    0.000 {built-in method torch.nn.functional.pad}
    10000    0.046    0.000    0.046    0.000 {method 'view' of 'torch._C._TensorBase' objects}
     2000    0.040    0.000    5.752    0.003 /vision/torchvision/prototype/transforms/_transform.py:66(forward)
     2000    0.038    0.000    0.138    0.000 /usr/lib/python3.8/traceback.py:321(extract)
     6000    0.036    0.000    0.036    0.000 {method 'sum' of 'torch._C._TensorBase' objects}
    10000    0.036    0.000    0.036    0.000 {built-in method posix.stat}
     6000    0.032    0.000    0.032    0.000 {built-in method torch.cumsum}
     2000    0.031    0.000    0.053    0.000 /usr/lib/python3.8/traceback.py:388(format)
     6000    0.028    0.000    0.028    0.000 {method 'clamp_' of 'torch._C._TensorBase' objects}
     2000    0.022    0.000    0.022    0.000 {built-in method torch.rand}
     2000    0.018    0.000    5.648    0.003 /vision/torchvision/prototype/transforms/functional/_color.py:209(equalize_image_tensor)
     2000    0.015    0.000    5.395    0.003 /vision/torchvision/prototype/transforms/functional/_color.py:223(<listcomp>)
    54000    0.013    0.000    0.033    0.000 /usr/lib/python3.8/traceback.py:285(line)
     2000    0.012    0.000    0.012    0.000 {method 'unbind' of 'torch._C._TensorBase' objects}
    36000    0.012    0.000    0.012    0.000 {method 'format' of 'str' objects}

@vfdev-5 vfdev-5 marked this pull request as ready for review October 11, 2022 09:59
datumbox
datumbox approved these changes Oct 11, 2022
View reviewed changes
Copy link
Copy Markdown
Contributor

@datumbox datumbox left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM, thanks! Feel free to merge on green CI.

torchvision/prototype/transforms/functional/_color.py
Comment on lines +202 to +203
lut.clamp_(0, 255)
lut = lut.to(torch.uint8)
Copy link
Copy Markdown
Contributor

@datumbox datumbox Oct 11, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you add a comment here explaining what we discussed offline in regards to why moving clamp and to here leads to a faster result?

@datumbox datumbox added module: transforms Perf For performance improvements prototype labels Oct 11, 2022
@vfdev-5
Copy link
Copy Markdown
Contributor Author

vfdev-5 commented Oct 11, 2022

There can be more improvement if we vectorize histogram computation with scatter_add_ (cc @lezcano).
As Vasilis is suggesting to go iteratively, let's merge this PR first and put other improvements in follow-up PRs.

@vfdev-5 vfdev-5 merged commit 11a2eed into pytorch:main Oct 11, 2022
@vfdev-5 vfdev-5 deleted the proto-small-optim-equalize branch October 11, 2022 21:47
facebook-github-bot pushed a commit that referenced this pull request Oct 17, 2022
@facebook-github-bot
[fbsync] [proto] Small improvement for tensor equalize op (#6738)
99cf56b 
Summary:
* [proto] Small improvement for tensor equalize op

* Fix code formatting

* Added a comment on the ops

Reviewed By: NicolasHug

Differential Revision: D40427464

fbshipit-source-id: f40623c83cebe269717151ae52f1fe9af47a3bde
@pmeier pmeier mentioned this pull request Oct 24, 2022
Closed
31 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

1 more reviewer

@datumbox datumbox datumbox approved these changes

Reviewers whose approvals may not affect merge requirements

Assignees

No one assigned

Labels

cla signed module: transforms Perf For performance improvements prototype

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@vfdev-5 @datumbox @facebook-github-bot