Add flexible bilinear upsampling aspect ratio by andrewgiessel · Pull Request #1279 · pytorch/pytorch

andrewgiessel · 2017-04-18T14:55:56Z

This PR addresses #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.

Some high level notes:

The largest changes refactor the Upsampling module (f2efeba) and functions (0760a41), moving everything from the base constructor down into the individual classes. I normally don't like removing that sort of abstraction, but I didn't see a good way to cleanly check the inputs for all cases at that higher level.
I changed scale_factor to be a pair (tuple) of non-negative ints instead of a single int, and edited two lines in the foward() of UpsamplingBilinear2d to reflect this. If a single number is passed in for scaling, it is duplicated in the tuple.
docstrings have been adjusted throughout
I added 2 tests. Not sure if this is sufficient, but given that nothing breaks, seems ok for now. As an aside, is there a way to run just one test?

Thanks, and input welcome. Tagging @soumith and @fmassa, who commented on the initial issue.

Rename the static library "libnccl_static.a" to disambiguate from the dynamic libraries.

Add a static library target "staticlib" to the Makefile.

Broadcast tuning Better checking of inputs Copy/reduce code simplification

Qualify nullptr_t with std::

This enables support for systems with more than 9 GPUs attached to a single PCIe root complex.

Functions vFetch and vStore are not found by ADL with clang, so they need to be declared before usage in ReduceCopy.

Fix compilation error when compiling with 'clang -x cuda'.

Adding tests for tuple scaling factors, even and uneven

torch/nn/_functions/thnn/upsampling.py

+        if size is None and scale_factor is None:
+            raise ValueError('either size or scale_factor should be defined')
+        if scale_factor is not None:
+            if not isinstance(scale_factor, (Integral, tuple)):


torch/nn/_functions/thnn/upsampling.py

+        if size is not None and not isinstance(size, tuple):
+            size = (size, size)
+        if scale_factor is not None and not isinstance(scale_factor, tuple):
+            scale_factor = (scale_factor, scale_factor)


torch/nn/_functions/thnn/upsampling.py

+    
    def __init__(self, size=None, scale_factor=None):
-        super(_UpsamplingBase, self).__init__()
+        super(UpsamplingNearest2d, self).__init__()


torch/nn/modules/upsampling.py

        [torch.FloatTensor of size 1x1x4x4]

    """
+    def __init__(self, size=None, scale_factor=None):


andrewgiessel · 2017-04-18T16:59:07Z

Thanks for the comments @apaszke!

also, the failed check is a linter error, I'll fix that right now.

edit: Bah- some of the inline comments are outdated now, sorry about that.

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

Fixes #1225. Signed-off-by: Edward Z. Yang <ezyang@fb.com>

The type checking is split across the parent and child classes

Adding tests for tuple scaling factors, even and uneven

The type checking is split across the parent and child classes

…b.com:andrewgiessel/pytorch into add-flexible-bilinear-upsampling-aspect-ratio

soumith · 2017-04-20T19:18:07Z

@andrewgiessel instead of rebasing on master, looks like you tried to merge via master. Now this branch has 64 commits :)

andrewgiessel · 2017-04-20T19:19:26Z

@fmassa UGH. I really tried... Suggestions? I can pop out the files that I've changed, delete, refork, and reopen the PR

andrewgiessel · 2017-04-20T19:22:28Z

Ya, I think I'm going to open a new PR and reference this one for posterity. 91 files changed? 🙄

andrewgiessel · 2017-04-20T19:28:28Z

let's try this again...

* Free output tensor on each pipeline stage for smaller memory footprint see: NVIDIA/Megatron-LM@057b086 * ref: NVIDIA/Megatron-LM@945ece9 * ref: NVIDIA/Megatron-LM@9a8b89a * remove position embedding group in destroy * pass deallocate_pipeline_outputs to backward_step * fix typo * missing deallocate_pipeline_outputs * fix typo: grad_ouptut -> grad_output * update tests * remove accessed todo * test with data parallel size of 2 if there's equal to or more than 8 gpus

sjeaugey and others added 26 commits October 7, 2016 12:42

Improved allreduce segmentation for small sizes

8bb06c9

NVML (libwrap) : import the needed definitions

bf7d151

Fix primitives function prototype

b2781d0

Bump to 1.3.1

534b9a1

Add Fortran bindings

5f2b32e

Add Copyright header to Fortran bindings source files

9c18468

Remove irrelevant output from ncclReduce Fortran tests

c2c5155

Add a static library target "staticlib" to the Makefile.

5765d60

Rename the static library "libnccl_static.a" to disambiguate from the dynamic libraries.

Merge pull request #54 from peterhj/peterhj-staticlib

ddddfba

Add a static library target "staticlib" to the Makefile.

Replace min BW by average BW in tests

1093821

1.3.2 release

34d2777

Broadcast tuning Better checking of inputs Copy/reduce code simplification

Adding missing file

648e9fb

Fix 1.3.2 compilation

2a974f5

Qualify nullptr_t with std::.

5eab428

Merge pull request #69 from cwhipkey/master

024d1e2

Qualify nullptr_t with std::

Fix crash in Reduce when non-root ranks have invalid recvbuff

8e1d6f9

Fix copy/paste typo in error message

c219a18

Only enable peer access for ring neighbors.

8996811

This enables support for systems with more than 9 GPUs attached to a single PCIe root complex.

Bumping version to 1.3.3

7fef264

Fix compilation error when compiling with 'clang -x cuda'.

8241cd7

Functions vFetch and vStore are not found by ADL with clang, so they need to be declared before usage in ReduceCopy.

Added Pascal nvcc flags, bumped version

649f04d

Merge pull request #78 from ilya-biryukov/master

ccfc456

Fix compilation error when compiling with 'clang -x cuda'.

Edits to nn/modules/upsampling.py

f2efeba

changes to upsampling functions

0760a41

Adding tests

8747c0d

Adding tests for tuple scaling factors, even and uneven

quick edit (debugging leftover)

c31ded0

apaszke reviewed Apr 18, 2017

View reviewed changes

Update ops for Sigmoid and Tanh

8a901c5

quick whitespace fix

c3dc907

Martin Raison and others added 21 commits April 18, 2017 12:46

make gradients contiguous in adagrad

1018b23

more operations and optimizations (hspmm, reorder, ...)

cd3bbc9

CPU hspmm + more efficient reorder

655c225

speed improvements, fix tests

701e631

Add some missing documentation for arguments.

ab77742

Signed-off-by: Edward Z. Yang <ezyang@fb.com>

Expose dilated convolutions.

34546f0

Fixes #1225. Signed-off-by: Edward Z. Yang <ezyang@fb.com>

Corrections in legacy modules. (#1286)

fc19473

Merge commit 'fcf4deac7d215f134ea25cd3def8b564b58b033c'

5a05109

Change the default algo for cuDNN conv forward to PRECOMP_GEMM (#1290)

934816c

Update README.md

c7d83a1

Fix loss.py docs (#1296)

6aa22be

Update functional.py (#1298)

a35f507

Fix argument typo in pad_packed_sequence docstring (#1300)

a8e6610

Restoring parent classes for bilinear sampling

73c0ea1

The type checking is split across the parent and child classes

Edits to nn/modules/upsampling.py

34b93bd

changes to upsampling functions

0ae4fd4

Adding tests

303249b

Adding tests for tuple scaling factors, even and uneven

quick edit (debugging leftover)

2b8b4de

quick whitespace fix

f57ed60

Restoring parent classes for bilinear sampling

e7c1d42

The type checking is split across the parent and child classes

Merge branch 'add-flexible-bilinear-upsampling-aspect-ratio' of githu…

b72e375

…b.com:andrewgiessel/pytorch into add-flexible-bilinear-upsampling-aspect-ratio

andrewgiessel closed this Apr 20, 2017

andrewgiessel mentioned this pull request Apr 21, 2017

Add flexible bilinear upsampling aspect ratio redux #1317

Merged

ezyang added the open source label Jun 24, 2019

Conversation

andrewgiessel commented Apr 18, 2017

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

andrewgiessel commented Apr 18, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

soumith commented Apr 20, 2017

Uh oh!

andrewgiessel commented Apr 20, 2017

Uh oh!

andrewgiessel commented Apr 20, 2017

Uh oh!

andrewgiessel commented Apr 20, 2017

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

17 participants

andrewgiessel commented Apr 18, 2017 •

edited

Loading