Dirac init compatibility with group convolutions #32825
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
💊 CircleCI build failures summary and remediationsAs of commit 94b539d:
Detailed failure analysisOne may explore the probable reasons each build failed interactively on the Dr. CI website. 🕵️ 3 new failures recognized by patternsThe following build failures do not appear to be due to upstream breakage:
|
assafshocher
force-pushed
the
dirac_groups
branch
from
a94f554 to
b2539e8
Compare
ezyang
added
the
triaged
|
@vincentqb do you think you would be able to review this? (No is a fine answer) |
vincentqb
suggested changes
Feb 7, 2020
There was a problem hiding this comment.
Couldn't you do the following instead?
x = torch.randn([1, 3, 3, 3])
print('input:\n', x)
conv_layer = torch.nn.Conv2d(3, 3, 3, padding=1, groups=1, bias=False)
torch.nn.init.dirac_(conv_layer.weight)
print('\noutput (before this PR):\n',conv_layer(x)) input:
tensor([[[[-0.4437, -0.2448, 1.4811],
[ 0.2103, 0.2234, 0.2437],
[-2.6258, 0.2899, -0.1680]],
[[-0.3531, -0.3139, -3.1256],
[ 0.5868, 0.3861, -0.0573],
[ 1.8914, 0.2475, 0.2421]],
[[-0.8827, 0.4264, -0.4944],
[ 1.8532, 0.0400, 1.4679],
[ 1.8510, 2.3384, 0.8307]]]])
output (before this PR):
tensor([[[[-0.4437, -0.2448, 1.4811],
[ 0.2103, 0.2234, 0.2437],
[-2.6258, 0.2899, -0.1680]],
[[-0.3531, -0.3139, -3.1256],
[ 0.5868, 0.3861, -0.0573],
[ 1.8914, 0.2475, 0.2421]],
[[-0.8827, 0.4264, -0.4944],
[ 1.8532, 0.0400, 1.4679],
[ 1.8510, 2.3384, 0.8307]]]], grad_fn=<MkldnnConvolutionBackward>)
I'm assuming the win is to do the identity by groups when out_channels is different from input_channels? But then this can be explicitly achieved by assigning with advanced indexing on a zero matrix of the final shape, no?
i.e. assign the sub tensor of
tensor([[[[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000]],
[[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000]],
[[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000]],
[[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000]]]], grad_fn=<MkldnnConvolutionBackward>)
the values of
tensor([[[[ 1.2205, -0.6608, 0.8640],
[-0.5464, 1.1288, 1.4726],
[-0.6693, 0.4000, -1.7613]],
[[-0.8760, -0.8814, -0.4705],
[ 0.6283, -0.5943, 0.6873],
[-0.6852, 1.4723, 0.3325]]]])
to get
tensor([[[[ 1.2205, -0.6608, 0.8640],
[-0.5464, 1.1288, 1.4726],
[-0.6693, 0.4000, -1.7613]],
[[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000]],
[[-0.8760, -0.8814, -0.4705],
[ 0.6283, -0.5943, 0.6873],
[-0.6852, 1.4723, 0.3325]],
[[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000]]]], grad_fn=<MkldnnConvolutionBackward>)
Moreover, if torch.nn.init.dirac_ supports groups, should torch.nn.init.* also?
|
@vincentqb |
vincentqb
approved these changes
Feb 12, 2020
There was a problem hiding this comment.
Your suggestions are valid if you wanted to do one time operation, but if you want to initialize a group conv layer, so changing back to groups=1 misses the point. Using indexing also. There is no problem copying a tensor but the point is initializing weights when using a group conv and do it consistently.
Moreover, currently what you get for a group conv initialized with Dirac doesn't make sense.
Reg init.*, Other initializations don't suffer from this inconsistency, since they don't need to create a structure that is channel aware like a dirc matrix, they simply fill with numbers all the channels without cross-dependencies so I think Dirac is the only one requires group support.
Sounds good, thanks!
There was a problem hiding this comment.
@vincentqb has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
|
@vincentqb |
None of the tests failing seem related. Landing. |
|
@vincentqb merged this pull request in 2c99ea8. |
ttumiel
pushed a commit
to ttumiel/pytorch
that referenced
this pull request
Mar 4, 2020
Summary:
Initializing weights of group-conv with init.dirac_, and applying, previously resulted in an output that makes no sense:
```
x = torch.randn([1, 3, 3, 3])
print('input:\n', x)
conv_layer = torch.nn.Conv2d(3, 3, 3, padding=1, groups=3, bias=False)
torch.nn.init.dirac_(conv_layer.weight.data)
print('\noutput (before this PR):\n',conv_layer(x))
input:
tensor([[[[ 0.5369, -1.1428, 0.1031],
[ 0.4638, -0.0854, -0.6553],
[ 0.8321, -2.5926, -0.3214]],
[[-0.2289, -0.0895, 0.4407],
[ 1.2309, -1.2096, -1.5216],
[-0.1798, 1.1694, 0.3469]],
[[ 0.1905, 0.8095, 0.5490],
[-0.4525, -0.4284, -0.1141],
[ 1.1857, -0.9246, -0.5119]]]])
output (before this PR):
tensor([[[[ 0.5369, -1.1428, 0.1031],
[ 0.4638, -0.0854, -0.6553],
[ 0.8321, -2.5926, -0.3214]],
[[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000]],
[[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000]]]], grad_fn=<MkldnnConvolutionBackward>)
````
This PR allows introducing groups to the initialization:
```
torch.nn.init.dirac_(conv_layer.weight.data, groups=3)
print('output (after this PR):\n', conv_layer(x))
output (after this PR):
tensor([[[[ 0.5369, -1.1428, 0.1031],
[ 0.4638, -0.0854, -0.6553],
[ 0.8321, -2.5926, -0.3214]],
[[-0.2289, -0.0895, 0.4407],
[ 1.2309, -1.2096, -1.5216],
[-0.1798, 1.1694, 0.3469]],
[[ 0.1905, 0.8095, 0.5490],
[-0.4525, -0.4284, -0.1141],
[ 1.1857, -0.9246, -0.5119]]]], grad_fn=<MkldnnConvolutionBackward>)
```
When out_channels is different than input_channels, it does the natural thing which is applying identity in each group separately:
```
x = torch.randn([1, 2, 3, 3])
print('input:\n', x)
conv_layer = torch.nn.Conv2d(2, 4, 3, padding=1, groups=2, bias=False)
torch.nn.init.dirac_(conv_layer.weight.data, groups=2)
print('\noutput:\n', conv_layer(x))
input:
tensor([[[[ 1.2205, -0.6608, 0.8640],
[-0.5464, 1.1288, 1.4726],
[-0.6693, 0.4000, -1.7613]],
[[-0.8760, -0.8814, -0.4705],
[ 0.6283, -0.5943, 0.6873],
[-0.6852, 1.4723, 0.3325]]]])
output:
tensor([[[[ 1.2205, -0.6608, 0.8640],
[-0.5464, 1.1288, 1.4726],
[-0.6693, 0.4000, -1.7613]],
[[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000]],
[[-0.8760, -0.8814, -0.4705],
[ 0.6283, -0.5943, 0.6873],
[-0.6852, 1.4723, 0.3325]],
[[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000],
[ 0.0000, 0.0000, 0.0000]]]], grad_fn=<MkldnnConvolutionBackward>)
```
Argument 'groups' defaults to 1 so it is backward compatible.
Tests are modified to include cases of with groups>1 but also contain groups=1 cases.
Pull Request resolved: pytorch#32825
Differential Revision: D19859926
Pulled By: vincentqb
fbshipit-source-id: 9dfdd24471ff14d79c442dfd28c1891aff812fdf
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Initializing weights of group-conv with init.dirac_, and applying, previously resulted in an output that makes no sense:
This PR allows introducing groups to the initialization:
When out_channels is different than input_channels, it does the natural thing which is applying identity in each group separately:
Argument 'groups' defaults to 1 so it is backward compatible.
Tests are modified to include cases of with groups>1 but also contain groups=1 cases.