Add flexible bilinear upsampling aspect ratio redux #1317

andrewgiessel · 2017-04-21T01:15:29Z

This PR addresses issue #1257, which aims to add non-coupled scaling factors for bilinear 2d upsampling. I've added two new tests, and currently everything passes. This is a relatively simple change which only touches python code.

Furthermore, this is is a re-do of PR #1279, wherein I messed up a rebase. There are some in-line comments there that might be worth skimming through, as well, but I tried to address most of the concerns raised there. In particular, I maintained the base class of all the upsampling methods.

Tests pass, on GPU and CPU.

tagging @apaszke @fmassa @soumith. Thanks in advance everyone.

ps: I'll probably need a pointer on how to best rebase to master, eventually ;)

torch/nn/_functions/thnn/upsampling.py

-        self.size = size
+        if scale_factor is not None and not isinstance(scale_factor, (Integral, tuple)):
+            raise ValueError('scale_factor must be of integer type or tuple of integer types')
+        self.size = _pair(size)


torch/nn/_functions/thnn/upsampling.py

+            # we have to be a tuple at this point
+            try:
+                assert len(self.scale_factor) == 2
+                for i in self.scale_factor:


torch/nn/_functions/thnn/upsampling.py

            self.output_size = (
-                input.size(2) * self.scale_factor,
-                input.size(3) * self.scale_factor,
+                input.size(2) * self.scale_factor[0],


torch/nn/modules/upsampling.py

@@ -110,5 +115,21 @@ class UpsamplingBilinear2d(_UpsamplingBase):

    """

+    def __init__(self, size=None, scale_factor=None):
+        super(UpsamplingBilinear2d, self).__init__(size, scale_factor)


torch/nn/modules/upsampling.py

-        if scale_factor is not None and not isinstance(scale_factor, Integral):
-            raise ValueError('scale_factor must be of integer type')
+        if scale_factor is not None and not isinstance(scale_factor, (Integral, tuple)):
+            raise ValueError('scale_factor must be of integer type or tuple of integer types')
        self.size = _pair(size)


torch/nn/_functions/thnn/upsampling.py

+
+        if self.scale_factor is not None:
+            self.scale_factor = _pair(self.scale_factor)
+            # we have to be a tuple at this point


Moved to functional.py

andrewgiessel · 2017-04-22T00:38:37Z

Thanks for the comments @apaszke! I think I made all the changes you asked for. Please, let me know if I didn't understand something (in particular I just c/p your __setstate__() function).

This allows for the base class to be used for upsampling routines other than 2d. I also renamed _check_bilinear_2d_scale_factor().

andrewgiessel · 2017-04-27T18:33:40Z

Please see the comment I made here on #1348 explaining the most recent commit.

soumith · 2017-05-03T12:46:35Z

thanks Andrew!

…8ffb52 (pytorch#11346) Summary: Pull Request resolved: pytorch#11346 Previous import was 1b09eb14c2c781fae078fa6b1c0390ba6fc0898c Included changes: - **[bff0b88](onnx/onnx@bff0b88)**: Add DynamicSlice experimental op (pytorch#1377) <James Reed> - **[91a7b8e](onnx/onnx@91a7b8e)**: statCoverage(model) (pytorch#1246) <Akshay Chalana> - **[36643c6](onnx/onnx@36643c6)**: fix the doc for softmax (pytorch#1374) <Lu Fang> - **[8c64acd](onnx/onnx@8c64acd)**: Silence usused result warning in ONNXIFI wrapper cleanup. Fix pytorch#1344 (pytorch#1371) <Marat Dukhan> - **[53b20f6](onnx/onnx@53b20f6)**: Add the ability to deprecate an OpSchema (pytorch#1317) <Ryan Hill> - **[8aec4e2](onnx/onnx@8aec4e2)**: [Anderspapitto patch] fix the shape inference for broadcasting (pytorch#1368) <Lu Fang> Reviewed By: jamesr66a Differential Revision: D9691533 fbshipit-source-id: 1a8c22262ae4946897e4be030d3f1cf3a3ad58b6

…8ffb52 (#11346) Summary: Pull Request resolved: #11346 Previous import was 1b09eb14c2c781fae078fa6b1c0390ba6fc0898c Included changes: - **[bff0b88](onnx/onnx@bff0b88)**: Add DynamicSlice experimental op (#1377) <James Reed> - **[91a7b8e](onnx/onnx@91a7b8e)**: statCoverage(model) (#1246) <Akshay Chalana> - **[36643c6](onnx/onnx@36643c6)**: fix the doc for softmax (#1374) <Lu Fang> - **[8c64acd](onnx/onnx@8c64acd)**: Silence usused result warning in ONNXIFI wrapper cleanup. Fix #1344 (#1371) <Marat Dukhan> - **[53b20f6](onnx/onnx@53b20f6)**: Add the ability to deprecate an OpSchema (#1317) <Ryan Hill> - **[8aec4e2](onnx/onnx@8aec4e2)**: [Anderspapitto patch] fix the shape inference for broadcasting (#1368) <Lu Fang> Reviewed By: jamesr66a Differential Revision: D9691533 fbshipit-source-id: 6aff6ce04ade37182e2ffe9bc83eb86846bc722d

…8ffb52 (pytorch#11346) Summary: Pull Request resolved: pytorch#11346 Previous import was 1b09eb14c2c781fae078fa6b1c0390ba6fc0898c Included changes: - **[bff0b88](onnx/onnx@bff0b88)**: Add DynamicSlice experimental op (pytorch#1377) <James Reed> - **[91a7b8e](onnx/onnx@91a7b8e)**: statCoverage(model) (pytorch#1246) <Akshay Chalana> - **[36643c6](onnx/onnx@36643c6)**: fix the doc for softmax (pytorch#1374) <Lu Fang> - **[8c64acd](onnx/onnx@8c64acd)**: Silence usused result warning in ONNXIFI wrapper cleanup. Fix pytorch#1344 (pytorch#1371) <Marat Dukhan> - **[53b20f6](onnx/onnx@53b20f6)**: Add the ability to deprecate an OpSchema (pytorch#1317) <Ryan Hill> - **[8aec4e2](onnx/onnx@8aec4e2)**: [Anderspapitto patch] fix the shape inference for broadcasting (pytorch#1368) <Lu Fang> Reviewed By: jamesr66a Differential Revision: D9691533 fbshipit-source-id: 6aff6ce04ade37182e2ffe9bc83eb86846bc722d

… and `fused_weight_gradient_mlp_cuda` is missing (pytorch#1317)

Andrew Giessel added 3 commits April 20, 2017 15:37

bilinear upsampling module

183314c

bilinear upsampling functions

3a058e1

bilinear upsampling tests

6611bb7

apaszke reviewed Apr 21, 2017

View reviewed changes

Andrew Giessel added 3 commits April 21, 2017 20:33

refactor out bilinear_scale_factor check

5752795

Moved to functional.py

checks for None in _pair() calls

435500f

adding __setstate__()

25b19f9

Andrew Giessel added 3 commits April 24, 2017 09:41

Adding Integral import

9513b24

actually return scale_factor after check

dbc26bd

sneak linting

c2f5f96

andrewgiessel mentioned this pull request Apr 25, 2017

Add upsample 3d, subsample 2d and 3d modules #1348

Closed

Move casting to tuples out of parent function

1d02692

This allows for the base class to be used for upsampling routines other than 2d. I also renamed _check_bilinear_2d_scale_factor().

soumith approved these changes May 3, 2017

View reviewed changes

soumith merged commit 2e7635b into pytorch:master May 3, 2017

Jiaming-Liu pushed a commit to Jiaming-Liu/pytorch that referenced this pull request May 18, 2017

Add flexible bilinear upsampling aspect ratio redux (pytorch#1317)

c5385ff

houseroad mentioned this pull request Sep 6, 2018

Automatic update of fbcode/onnx to bff0b8835870c7df7762ef43498d000d2d8ffb52 #11346

Closed

ezyang added the open source label Jun 24, 2019

hubertlu-tw pushed a commit to hubertlu-tw/pytorch that referenced this pull request Nov 1, 2022

[transformer] Warn only when gradient_accumulation_fusion is True…

7950a82

… and `fused_weight_gradient_mlp_cuda` is missing (pytorch#1317)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add flexible bilinear upsampling aspect ratio redux #1317

Add flexible bilinear upsampling aspect ratio redux #1317

andrewgiessel commented Apr 21, 2017

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

andrewgiessel commented Apr 22, 2017

andrewgiessel commented Apr 27, 2017

soumith commented May 3, 2017

Add flexible bilinear upsampling aspect ratio redux #1317

Add flexible bilinear upsampling aspect ratio redux #1317

Conversation

andrewgiessel commented Apr 21, 2017

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

andrewgiessel commented Apr 22, 2017

andrewgiessel commented Apr 27, 2017

soumith commented May 3, 2017