instance norm and reflection padding #7938

zhanghang1989 · 2017-09-18T19:56:27Z

Re-create the PR for instance norm and reflection padding.

szha · 2017-09-18T22:48:35Z

Could you fix lint? The errors can be obtained by running make pylint

piiswrong · 2017-09-20T19:58:24Z

python/mxnet/gluon/nn/basic_layers.py

+    Output shape:
+        Same shape as input.
+
+       This implementation is based on paper:


fix format. see dropout

piiswrong · 2017-09-20T19:58:34Z

python/mxnet/gluon/nn/conv_layers.py

+
+    Parameters
+    ----------
+        pad_width int: the size of the padding. If is int, uses the same


if not int what else can it be?

piiswrong · 2017-09-20T19:58:43Z

python/mxnet/gluon/nn/conv_layers.py

+    >>> input = mx.nd.random_normal(shape=(16, 3, 224, 224))
+    >>> output = m(input)
+    """
+    def __init__(self, pad_width=0, **kwargs):


pad_width -> padding

piiswrong · 2017-09-20T19:59:33Z

python/mxnet/gluon/nn/conv_layers.py

+        pad_width int: the size of the padding. If is int, uses the same
+            padding in all boundaries. 
+
+    Shape:


piiswrong · 2017-09-20T20:00:05Z

python/mxnet/gluon/nn/conv_layers.py

+    Examples
+    --------
+    >>> m = nn.ReflectionPad(3)
+    >>> input = mx.nd.random_normal(shape=(16, 3, 224, 224))


mx.nd.random.normal

piiswrong · 2017-09-20T20:00:14Z

python/mxnet/gluon/nn/conv_layers.py

+
+    Examples
+    --------
+    >>> m = nn.ReflectionPad(3)


m -> layer

ReflectionPad -> ReflectionPad2D

piiswrong · 2017-09-20T20:01:28Z

python/mxnet/gluon/nn/conv_layers.py

+        self._pad_width = pad_width
+
+    def forward(self, x):
+        return F.pad(x, mode='reflect', pad_width=self._pad_width)


shouldn't pad_width be a tuple?

piiswrong · 2017-09-20T20:01:34Z

Please add tests.

piiswrong · 2017-12-12T21:55:49Z

@szha The original author seems to have disappeared. Would you like to take over?

szha · 2017-12-12T22:11:17Z

Actually @zhanghang1989 will join us soon 😄

zhanghang1989 · 2017-12-19T04:04:45Z

@piiswrong @szha Hi guys, sorry for the delay. I just made some changes.

piiswrong · 2018-01-03T19:55:22Z

see the other ops for doc format.

Also please add test cases

zhanghang1989 · 2018-01-03T21:09:26Z

Thanks @piiswrong! I made some changes to the doc.

eric-haibin-lin

Is there no unit test?

zhanghang1989 · 2018-01-12T19:22:18Z

I will add one unit test soon at unittest/test_gluon.py . Thanks!

szha

You need to add the InstanceNorm class in __all__ here https://github.com/apache/incubator-mxnet/pull/7938/files#diff-06285baf97b6456dd23bdb99f6e239b6R23

szha · 2018-01-31T18:11:41Z

please fix lint

szha · 2018-01-31T19:07:37Z

python/mxnet/gluon/nn/conv_layers.py

+        self._padding = padding
+
+    def hybrid_forward(self, F, x):
+        return F.pad(x, mode='reflect', padding=self._padding)


This isn't tested yet. Also, it looks like a very straightforward one-liner. Should we just ask users to use hybrid lambda for this?

This simple API makes user easier to use. The Reflectance padding is popularly used in low level vision, such as GANs, super resolution and style transfer. I think it make sense to provide an API for it. What's your thought?

http://pytorch.org/docs/0.3.0/nn.html?highlight=padding#reflectionpad2d

I think we should have this.

piiswrong · 2018-01-31T19:33:28Z

python/mxnet/gluon/nn/conv_layers.py

+
+    Parameters
+    ----------
+        `padding` is a tuple of integer padding widths for each axis of the format


Please fix the documentation and add tests

piiswrong · 2018-01-31T19:34:12Z

python/mxnet/gluon/nn/conv_layers.py

+    """
+    def __init__(self, padding=0, **kwargs):
+        super(ReflectionPad2D, self).__init__(**kwargs)
+        self._padding = padding


Are you sure this works correctly when padding is a number instead of a tuple?

just updated the integer type check and fix docs

szha

Add the classes to docs/api/python/gluon/nn.md so that doc is available.

szha · 2018-02-02T05:25:20Z

python/mxnet/gluon/nn/basic_layers.py

+        super(InstanceNorm, self).__init__(**kwargs)
+        self._kwargs = {'eps': epsilon}
+        if in_channels != 0:
+            self.in_channels = in_channels


This doesn't seem needed in forward. The reference in __repr__ can be changed to use the shape from weight instead.

szha · 2018-02-02T05:26:07Z

python/mxnet/gluon/nn/basic_layers.py

+    def __repr__(self):
+        s = '{name}({content}'
+        if hasattr(self, 'in_channels'):
+            s += ', in_channels={0}'.format(self.in_channels)


load size from weight instead.

szha · 2018-02-02T05:28:25Z

python/mxnet/gluon/nn/conv_layers.py

+    padding: int or a tuple of int
+        An integer padding width for height and weight or a tuple of
+        integers padding widths for each axis of the format ``(before_1, after_1, ... ,
+        before_N, after_N)``. The `padding` should be of length ``2*N`` where ``N``


The document describes something general but the class is specific to 2D. Maybe describe the argument that's specific to 2D case here.

If pass into a tuple, it works for different dimensions. Should we remove this?

Since the class name says 2D, it should support 2D.

You can do something similar to the conv blocks, with a common abstract class for the forward logic, and the actual 1D/2D/3D classes for specific implementation.

Agree. Just made the changes. Will refactor when adding 1D/3D

szha · 2018-02-03T01:53:15Z

python/mxnet/gluon/nn/basic_layers.py

@@ -555,8 +553,9 @@ def hybrid_forward(self, F, x, gamma, beta):

    def __repr__(self):
        s = '{name}({content}'
+        in_channels = self.gamma.shape[0]
        if hasattr(self, 'in_channels'):


remove condition

piiswrong · 2018-02-13T19:36:01Z

@zhanghang1989 @szha
The instance norm operator needs an axis argument to be consistent with batchnorm

reflectionpad2d's input/output shape doc has wrong format

* instance norm and reflection padding * r prefix * indent and space * fix docs * change docs * spacing * typo * hybrid forward * spcaing * add test for instance norm * fix typo * add to __all__ * rm white space * integer value * add test * make line short * rm white space * add docs ref * fix docs * RFpad2D docs * read shape from weight * rm condition

instance norm and reflection padding

fdcb42c

zhanghang1989 requested review from mli and piiswrong as code owners September 18, 2017 19:56

zhanghang1989 added 2 commits September 19, 2017 12:33

r prefix

0e1ca0b

indent and space

cc547e9

piiswrong reviewed Sep 20, 2017

View reviewed changes

fix docs

a491219

zhanghang1989 added 2 commits January 3, 2018 13:02

Merge github.com:apache/incubator-mxnet into patch-1

847779e

change docs

810e138

zhanghang1989 added 4 commits January 3, 2018 13:21

spacing

c9abee4

typo

647d4be

hybrid forward

cbd19fc

spcaing

6e658de

eric-haibin-lin reviewed Jan 12, 2018

View reviewed changes

add test for instance norm

3bf128b

zhanghang1989 requested a review from szha as a code owner January 29, 2018 23:56

zhanghang1989 added 2 commits January 29, 2018 15:56

Merge github.com:apache/incubator-mxnet into patch-1

c06b250

fix typo

b7a5a6b

szha reviewed Jan 31, 2018

View reviewed changes

add to __all__

d6b373d

rm white space

ebcaf66

szha reviewed Jan 31, 2018

View reviewed changes

piiswrong reviewed Jan 31, 2018

View reviewed changes

zhanghang1989 added 4 commits January 31, 2018 11:55

integer value

d119f31

add test

b416765

make line short

4838ed1

rm white space

de390aa

szha reviewed Feb 1, 2018

View reviewed changes

zhanghang1989 added 2 commits February 1, 2018 14:53

add docs ref

870fe25

fix docs

ae585bc

szha reviewed Feb 2, 2018

View reviewed changes

zhanghang1989 added 2 commits February 2, 2018 10:57

RFpad2D docs

fe1d008

read shape from weight

fd29d82

szha reviewed Feb 3, 2018

View reviewed changes

rm condition

fbd2b3c

szha merged commit d452162 into apache:master Feb 3, 2018

zhanghang1989 deleted the patch-1 branch February 6, 2018 23:28

instance norm and reflection padding #7938

instance norm and reflection padding #7938

Conversation

zhanghang1989 commented Sep 18, 2017

szha commented Sep 18, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

piiswrong Sep 20, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

piiswrong commented Sep 20, 2017

piiswrong commented Dec 12, 2017

szha commented Dec 12, 2017

zhanghang1989 commented Dec 19, 2017

piiswrong commented Jan 3, 2018

zhanghang1989 commented Jan 3, 2018

eric-haibin-lin left a comment

Choose a reason for hiding this comment

zhanghang1989 commented Jan 12, 2018

szha left a comment • edited Loading

Choose a reason for hiding this comment

szha commented Jan 31, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

szha left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

piiswrong commented Feb 13, 2018

piiswrong Sep 20, 2017 •

edited

Loading

szha left a comment •

edited

Loading