Add AffineTransformation #793

ekagra-ranjan · 2019-03-11T07:25:43Z

In reference to #569 added AffineTransformation to superseed LinearTransformation.

update before transform

update 9/03/19

11/03/19

Add Affinetransformation to superseed LinearTransformation

fmassa

Hi,

Thanks for the PR!

I have one comment about LinearTransformation that I think should be addressed before this can be merged, otherwise this looks good, thanks!

fmassa · 2019-03-11T10:19:57Z

torchvision/transforms/transforms.py

+    def __init__(self, transformation_matrix):
+        warnings.warn("The use of the transforms.LinearTransformation transform is deprecated, " +
+                      "please use transforms.AffineTransformation instead.")
+        super(LinearTransformation, self).__init__(transformation_matrix)


I suppose you want to pass a mean_vector filled with zeros here?

What would you prefer for the dimension of mean_vector: 1xD or Dx1 given that transformation_matrix is DxD?

According to the implementation of LinearTransformation I think it should be 1 x D.

I feel that the doc of the application of LinearTransformation about whitening is broken. The doc only shows the 1st line of the paragraph in the application block and the rest outside it. How do we write a multi-line bullet point? Do we replace - with ---?

@ekagra-ranjan the dimension of the tensor doesn't matter, you can pass a D tensor and it will be fine, given that broadcasting will happen and the tensor will be filled with all zeros anyway.

About the doc, I'm not sure about how to properly format things. I'd try compiling the documentation locally and checking it visually

torchvision/transforms/transforms.py

11/03/19 6:41pm

codecov-io · 2019-03-11T20:11:45Z

Codecov Report

Merging #793 into master will increase coverage by 0.03%.
The diff coverage is 66.66%.

@@            Coverage Diff             @@
##           master     #793      +/-   ##
==========================================
+ Coverage    38.1%   38.13%   +0.03%     
==========================================
  Files          32       32              
  Lines        3128     3136       +8     
  Branches      488      489       +1     
==========================================
+ Hits         1192     1196       +4     
- Misses       1857     1860       +3     
- Partials       79       80       +1

Impacted Files	Coverage Δ
torchvision/transforms/transforms.py	`82.79% <66.66%> (-0.62%)`	⬇️
torchvision/models/resnet.py	`17.29% <0%> (ø)`	⬆️

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 8257aff...5883ef0. Read the comment docs.

torchvision/transforms/transforms.py

ekagra-ranjan · 2019-03-20T15:19:51Z

@fmassa Is there anything else that you want me to do with this PR?

ekagra-ranjan · 2019-03-24T19:28:25Z

@fmassa Are the changes fine?

fmassa

I have a couple more comments, after that I believe this will be good to go

fmassa · 2019-03-25T10:17:41Z

torchvision/transforms/transforms.py

        if transformation_matrix.size(0) != transformation_matrix.size(1):
            raise ValueError("transformation_matrix should be square. Got " +
                             "[{} x {}] rectangular matrix.".format(*transformation_matrix.size()))
+
+        if mean_vector.size(1) != transformation_matrix.size(0):


this is raise an error if the user tries to use LinearTransformation, because the mean vector is 1d.

I really think that you should just make mean_vector here be 1d as well, and change it in the documentation.

Also, can you keep the tests for the LinearTransformation?

fmassa · 2019-03-25T10:18:26Z

torchvision/transforms/transforms.py

        format_string += (str(self.transformation_matrix.numpy().tolist()) + ')')
+        format_string += (", (mean_vector=" + str(self.mean_vector.numpy().tolist()) + ')')


no need to convert to numpy here, tolist() also exists for torch tensors

…at_string

fmassa

Thanks!

yanfengliu · 2019-04-04T15:44:09Z

Sorry if I misunderstood this, but I thought affine transformation means rotate, scale, shift, and shear, as shown in torchvision.transforms.functional.affine(img, angle, translate, scale, shear, resample=0, fillcolor=None)?

fmassa · 2019-04-08T19:29:09Z

Hum, there seems to be indeed a problem with our naming, which will indeed make this very confusing

Given that this is not in the 0.2.2 release, I'll rename this function with another name.
Any advices on the naming?

ekagra-ranjan · 2019-04-08T20:43:05Z

I have 2 suggestions for the name:

We can let the name remain LinearTransformation as the new transformation is still a linear operation ( (X-mean)/std = AX - B, for some suitable A and B).
We can rename it as PixelNormalization since the transformation is basically is normalizing the values at each pixel in the image with pixel specific mean and standard deviation.

fmassa · 2019-04-09T08:20:31Z

@ekagra-ranjan I agree with leaving the name LinearTransformation.
Can you send a PR with this change?

ekagra-ranjan · 2019-04-09T13:03:34Z

Sure.

ekagra-ranjan added 5 commits February 14, 2019 22:34

Merge pull request #1 from pytorch/master

154cadb

update before transform

Merge pull request #2 from pytorch/master

e964780

update 9/03/19

Merge pull request #3 from pytorch/master

103b25a

11/03/19

Add Affinetransformation

c99eb98

Add Affinetransformation to superseed LinearTransformation

Add test

48d9002

ekagra-ranjan mentioned this pull request Mar 11, 2019

Add AffineTransformation #792

Closed

fmassa requested changes Mar 11, 2019

View reviewed changes

Add zero mean_vector in LinearTransformation and improved docs

1d057de

fmassa reviewed Mar 11, 2019

View reviewed changes

torchvision/transforms/transforms.py Outdated Show resolved Hide resolved

ekagra-ranjan added 18 commits March 11, 2019 18:39

Merge pull request #4 from pytorch/master

f538d46

11/03/19 6:41pm

update

4d24d94

minor fix

be21da5

minor fix2

2c57a56

fixed flake8

c8d5711

fix flake8

9462789

fixed transpose syntax

c257605

fixed shape of mean_vector in test

defa1da

fixed test

9b8d07b

print est cov and mean

cf7afcd

fixed flake8

a1ee3d1

debug

881eb41

reduce num_samples

f498aa5

debug

606f426

fixed num_features

6b5bab0

fixed rtol for cov

0238549

fix __repr__

12e6685

Update transforms.py

121c1cf

ekagra-ranjan added 2 commits March 12, 2019 01:41

Update test_transforms.py

5b7a445

Update transforms.py

c2d7627

ekagra-ranjan added 3 commits March 12, 2019 10:16

fix flake8

f524dfa

Update transforms.py

18bd5fc

Update transforms.py

e3735dc

fmassa reviewed Mar 12, 2019

View reviewed changes

torchvision/transforms/transforms.py Outdated Show resolved Hide resolved

ekagra-ranjan added 2 commits March 12, 2019 11:35

Update transforms.py

d5dcf4a

Update transforms.py

c727945

ekagra-ranjan changed the title ~~Patch AffineTransformation~~ Add AffineTransformation Mar 16, 2019

fmassa requested changes Mar 25, 2019

View reviewed changes

ekagra-ranjan added 3 commits March 25, 2019 18:02

Changed dim of mean_vector to 1D, doc and removed .numpy () from form…

9907dee

…at_string

Restore test_linear_transformation()

04c70ca

Update test_transforms.py

5883ef0

fmassa approved these changes Mar 25, 2019

View reviewed changes

fmassa merged commit c88d7fb into pytorch:master Mar 25, 2019

fmassa mentioned this pull request Mar 25, 2019

Whitening transform that supports mean subtraction #569

Closed

ekagra-ranjan mentioned this pull request Apr 9, 2019

Change AffineTransformation back to LinearTransformation #843

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add AffineTransformation #793

Add AffineTransformation #793

ekagra-ranjan commented Mar 11, 2019

fmassa left a comment

fmassa Mar 11, 2019

ekagra-ranjan Mar 11, 2019

ekagra-ranjan Mar 11, 2019

fmassa Mar 11, 2019

codecov-io commented Mar 11, 2019 •

edited

ekagra-ranjan commented Mar 20, 2019

ekagra-ranjan commented Mar 24, 2019

fmassa left a comment

fmassa Mar 25, 2019

fmassa Mar 25, 2019

fmassa left a comment

yanfengliu commented Apr 4, 2019

fmassa commented Apr 8, 2019

ekagra-ranjan commented Apr 8, 2019

fmassa commented Apr 9, 2019

ekagra-ranjan commented Apr 9, 2019

		format_string += (str(self.transformation_matrix.numpy().tolist()) + ')')
		format_string += (", (mean_vector=" + str(self.mean_vector.numpy().tolist()) + ')')

Add AffineTransformation #793

Add AffineTransformation #793

Conversation

ekagra-ranjan commented Mar 11, 2019

fmassa left a comment

Choose a reason for hiding this comment

fmassa Mar 11, 2019

Choose a reason for hiding this comment

ekagra-ranjan Mar 11, 2019

Choose a reason for hiding this comment

ekagra-ranjan Mar 11, 2019

Choose a reason for hiding this comment

fmassa Mar 11, 2019

Choose a reason for hiding this comment

codecov-io commented Mar 11, 2019 • edited

Codecov Report

ekagra-ranjan commented Mar 20, 2019

ekagra-ranjan commented Mar 24, 2019

fmassa left a comment

Choose a reason for hiding this comment

fmassa Mar 25, 2019

Choose a reason for hiding this comment

fmassa Mar 25, 2019

Choose a reason for hiding this comment

fmassa left a comment

Choose a reason for hiding this comment

yanfengliu commented Apr 4, 2019

fmassa commented Apr 8, 2019

ekagra-ranjan commented Apr 8, 2019

fmassa commented Apr 9, 2019

ekagra-ranjan commented Apr 9, 2019

codecov-io commented Mar 11, 2019 •

edited