This repository has been archived by the owner on Nov 17, 2023. It is now read-only.

[MXNet-1211] Factor and "Like" modes in BilinearResize2D operator #13226

Merged

wkcn merged 9 commits into apache:master from lobanov-m:bilinear-resize-scale-like

May 5, 2019

Contributor

lobanov-m commented Nov 12, 2018 •

edited

Loading

Description

This PR adds few options to BilinearResize2D operator. This options are needed for Fully-Convolutional realization of nets, that uses bilinear resize. The added modes are "odd_scale", "like" and few options to change height and width of tensor to even or odd: "to_even_down", "to_even_up", "to_odd_down", "to_odd_up".
Such transformations couldn't be done in other ways because when we use symbolic models we can't get tensors shapes from Python or any other frontend.
With these mods it is is possible to realize segmentation network ICNet that could work with input of any size.

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

Changes are complete (i.e. I finished coding on this PR)
For user-facing API changes, API doc string has been updated.
To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

"Odd-scale" is for getting odd height and width, for example for pooling, which is recommended by authors of DeepLab and ICNet. So if we have tensor with shape (n, c, 65, 33) we can resize it to (n, c, 33, 17) with parameters (scale_height=0.5, scale_width=0.5, mode='odd_scale') and back with parameters (scale_height=2, scale_width=2, mode='odd_scale')
"Like" is also useful mode which could be used in decoders of segmentation networks. It resizes first input to the shape of second input.
"to_even_..." and "to_odd_..." modes changes height and width to even and odd value respectively adding or subtracting 1 from h and w if required.
Changes in visualization.py are needed for normal graph visualization when using mode "Like". In this mode there are two input nodes, but actually one of them is used only for getting desired shape. Information about it in graph isn't important but graph becomes too complicated. So this change just deletes dependence from "like"-node.

Comments

The changes are almost backward-compatible: since we added mode "like", the operator takes two inputs, so in other modes second positional argument in Python (or other) should be None, and height and width became third and forth positional arguments respectively. So, some code, which used height and width as positional argument will not work.

lobanov-m requested review from anirudh2290 and szha as code owners

November 12, 2018 16:56

lobanov-m changed the title ~~Bilinear resize scale like~~ Factor and "Like" modes in BilinearResize2D operator

Member

anirudhacharya commented Nov 13, 2018

@mxnet-label-bot add [pr-awaiting-review]

marcoabreu added the pr-awaiting-review label

lobanov-m changed the title ~~Factor and "Like" modes in BilinearResize2D operator~~ [MXNet-1211] Factor and "Like" modes in BilinearResize2D operator

Contributor

stu1130 commented Nov 21, 2018

@apeforest @samskalicky @yuxihu @ChaiBapchya could you please take a look at this?

anirudhacharya reviewed

View reviewed changes

Member

anirudhacharya left a comment

Can you please add tests for this operator?

src/operator/contrib/bilinear_resize-inl.h Outdated

- .describe("output height (required)");
- DMLC_DECLARE_FIELD(width).set_range(1, 10000)
- .describe("output width (required)");
+ DMLC_DECLARE_FIELD(height).set_range(0, 10000)

Member

anirudhacharya Nov 21, 2018

why change the range from 1 to 0. How does 0 size make sense for resizing?

Contributor Author

lobanov-m Nov 26, 2018 •

edited

Loading

It is required for using "scale"-mode. F.e. you have tensor with shape (1, 3, 100, 100), so you can do BilinearResize2D(tensor, height=0.5, width=0.5, mode='scale') and get tensor of shape (1, 3, 50, 50)

src/operator/contrib/bilinear_resize-inl.h Outdated

- DMLC_DECLARE_FIELD(width).set_range(1, 10000)
- .describe("output width (required)");
+ DMLC_DECLARE_FIELD(height).set_range(0, 10000)
+ .set_default(-1)

Member

anirudhacharya Nov 21, 2018

height and width are currently required parameters. Why set a default value for them? Also how does setting default=-1 make sense when the range for the field is (0,1000)

Contributor Author

lobanov-m Nov 26, 2018 •

edited

Loading

It is not required parameter, when using mode 'like'. In this mode we can pass two tensors and the first tensor would be resized to second tensor's size. But for other modes it is still required, so default value is inappropriate and should raise error.

src/operator/contrib/bilinear_resize-inl.h Outdated

+ .describe(R"(output height if mode is "size" or scale on which input height should be multiplied if mode is
+ "scale" or "odd_scale")");
+ DMLC_DECLARE_FIELD(width).set_range(0, 10000)
+ .set_default(-1)

Member

anirudhacharya Nov 21, 2018

same as above.

src/operator/contrib/bilinear_resize-inl.h Outdated

+ .add_enum("like", bilinear_resize::like)
+ .add_enum("to_even", bilinear_resize::to_even)
+ .set_default(bilinear_resize::size)
+ .describe(R"(resizing mode. "size" - resize to distinct size; "scale" - original height and width are multiplied

Member

anirudhacharya Nov 21, 2018

in scale mode what is the "appropriate scale"? I dont see that defined in the params.

Contributor Author

lobanov-m Nov 26, 2018

I think any. If we have original tensor with shape (1, 3, 10000, 10000) and resize it with height=0.001 and width=0.001 in "scale" mode we will get tensor with shape (1, 3, 10, 10). And back we can resize the result tensor in scale mode with height=100 and width=100 and get tensor of original size. I'll try to clarify this in description.

Contributor Author

lobanov-m Nov 26, 2018

Probably I used wrong term. Maybe it would be better if I change "scale" to "factor".

src/operator/contrib/bilinear_resize-inl.h Outdated

+ .set_default(bilinear_resize::size)
+ .describe(R"(resizing mode. "size" - resize to distinct size; "scale" - original height and width are multiplied
+ by appropriate scale; "odd_scale" - the same, but result height and width are odd value
+ (if h%2==0 then h=h+1 and the same for width); "like" - resize first input to the height

Member

anirudhacharya Nov 21, 2018

it would be good to clarify what h is here? original image height or the height parameter in the parameter list.

Contributor Author

lobanov-m Nov 26, 2018

Yes, you are right, I'll do this, thank you.

src/operator/contrib/bilinear_resize-inl.h

+ CHECK_EQ(inputs.size(), 1);
+ bool modeLike = param.mode == bilinear_resize::like;
+ size_t expected = modeLike ? 2 : 1;
+ CHECK_EQ(outputs.size(), expected);

Member

anirudhacharya Nov 21, 2018

so there will be 2 outputs in like mode? what is the second output? if it does have two outputs, it should be specified in the param description.

Contributor Author

lobanov-m Nov 26, 2018 •

edited

Loading

There are two outputs of the backward function. The operator still has one output. In "like" mod we pass two input tensors to resize one to the size of second, so the backward function should return gradients to both tensors. Actually the second tensor, from which we get result size, should get zero gradients from the output of this operator, because it is needed only to get it's shape. It is realized in cc and cu files.

lobanov-m requested review from gigasquid, marcoabreu, nswamy and yzhliu as code owners

November 26, 2018 21:12

Contributor Author

lobanov-m commented Nov 26, 2018

I'm realy sorry for force push, I get confused with rebase and commited all master to this branch, after that tried to roll back and finally erased my first commits. :(

anirudhacharya reviewed

View reviewed changes

tests/python/unittest/test_operator.py

@@ @@ -6527,12 +6527,64 @@ def check_bilinear_resize_op(shape, height, width): @@
  x = mx.nd.random.uniform(shape=shape)
  y = mx.nd.contrib.BilinearResize2D(x, height=height, width=width)
  assert_almost_equal(y.asnumpy(), py_bilinear_resize(x.asnumpy(), height, width))
+ def check_bilinear_resize_modes_op(shape, height_factor=None, width_factor=None, shape_1=None, mode=None):
+ x = mx.nd.random.uniform(shape=shape)

Member

anirudhacharya Nov 27, 2018

unit tests for operators should cover forward pass and backward pass. See here - https://mxnet.incubator.apache.org/faq/add_op_in_backend.html#unit-test

Contributor Author

lobanov-m Nov 30, 2018

Done :)

ChaiBapchya suggested changes

View reviewed changes

Contributor

ChaiBapchya left a comment

Looks like it is failing on 2 different validation CI checks
1 looks like uncovered a flaky test
other has build failure which needs your attention. Please review and make requisite changes. Thanks.

ChaiBapchya mentioned this pull request

Flaky test test_random.test_negative_binomial_generator #13506

Closed

Contributor Author

lobanov-m commented Dec 3, 2018

Thank you, @ChaiBapchya.
As I understood the build has failed because of the timeout.
Branches [GPU: MKLDNN ] and [GPU: MKLDNN_CUDNNOFF] has failed.
In pipeline there are lines:

[GPU: MKLDNN] Running on mxnetlinux-cpu_x4398dm2a1 in /home/jenkins_slave/workspace/alidation_unix-gpu_PR-13226-7WOHJMMJA2UQ45SCSTCQK6UH3VENFWCLRGEEC4X3CRC25ELVL4IQ   
[Pipeline] [GPU: MKLDNN] {    
[Pipeline] [GPU: MKLDNN] ws    
[GPU: MKLDNN] Running in /home/jenkins_slave/workspace/build-mkldnn-gpu    
[Pipeline] [GPU: MKLDNN] {    
[Pipeline] [GPU: MKLDNN] timeout    
[GPU: MKLDNN] Timeout set to expire in 2 hr 0 min    
[Pipeline] [GPU: MKLDNN] {

And in the log file of build there is no error and it starts at 19:30 and fails at 21:30:

[build-mkldnn-gpu] Running shell script
+ ci/build.py --docker-registry mxnetci --platform ubuntu_build_cuda --docker-build-retries 3 --shm-size 500m /work/runtime_functions.sh build_ubuntu_gpu_mkldnn
build.py: 2018-11-30 19:30:49,272Z INFO MXNet container based build tool.
...
Sending interrupt signal to process
build.py: 2018-11-30 21:30:29,503Z WARNING Signal 15 received, cleaning up...
build.py: 2018-11-30 21:30:29,503Z WARNING Cleaning up containers
Terminated
script returned exit code 143

The same for [GPU: MKLDNN_CUDNNOFF]. No error in log, build script lasts for 1h 59m 55s and receives interrupt signal. Jenkins build description.

What should I do with this problem?

Contributor

azai91 commented Dec 4, 2018

@lobanov-m this was around the time we were changing our CI. can you retrigger the build (push an empty commit).

Contributor Author

lobanov-m commented Dec 11, 2018

Last mentioned problem with pull request have been resolved. Is anything else should be changed?

Contributor

sandeep-krishnamurthy commented Dec 29, 2018

@anirudhacharya @azai91 - ping for review :-)

anirudhacharya approved these changes

View reviewed changes

Member

anirudhacharya left a comment

LGTM

Contributor

sandeep-krishnamurthy commented Jan 26, 2019

@apeforest - Can you please help review this PR?
@lobanov-m - Can you please rebase? Thanks for your contributions

Contributor

vandanavk commented Feb 5, 2019

@mxnet-label-bot update [pr-awaiting-response]

marcoabreu added pr-awaiting-response and removed pr-awaiting-review labels

Contributor Author

lobanov-m commented Feb 14, 2019

Rebase done.

Contributor Author

lobanov-m commented Feb 19, 2019

@mxnet-label-bot update [pr-awaiting-review]

marcoabreu added pr-awaiting-review and removed pr-awaiting-response labels

Member

anirudhacharya commented Mar 2, 2019

@mxnet-label-bot add [pr-awaiting-merge]

@lobanov-m please rebase your PR to resolve conflicts.

Contributor Author

lobanov-m commented Mar 25, 2019

@mxnet-label-bot add[pr-awaiting-review]

marcoabreu added the pr-awaiting-review label

Contributor Author

lobanov-m commented Mar 25, 2019

@mxnet-label-bot remove[pr-awaiting-response]

marcoabreu removed the pr-awaiting-response label

Contributor Author

lobanov-m commented Mar 25, 2019

@pinaraws rebase done

Contributor Author

lobanov-m commented Mar 25, 2019

Built failed because of

test_model.R:129: error: Fine-tune
cannot open URL 'http:https://data.dmlc.ml/models/imagenet/inception-bn/Inception-BN-0126.params'
1: GetInception() at R-package/tests/testthat/test_model.R:129

and the second error:

--2019-03-25 08:58:30--  (try: 2)  http:https://data.dmlc.ml/mxnet/models/imagenet/inception-bn.tar.gz
Connecting to data.dmlc.ml (data.dmlc.ml)|54.208.175.7|:80... failed: Connection timed out.
Retrying.

Contributor Author

lobanov-m commented Mar 26, 2019 •

edited

Loading

Error in build doesn't connected with pull request:

Rscript -e "library(devtools); library(methods); options(repos=c(CRAN='https://cloud.r-project.org/')); install_deps(pkg='R-package', dependencies = TRUE)"
Error in library(devtools) : there is no package called 'devtools'

Contributor

piyushghai commented Apr 5, 2019

@anirudhacharya Are all your concerns addressed on this PR ?
Is this good to merge ?

Member

anirudhacharya commented Apr 5, 2019

@piyushghai yes, my concerns are addressed, please see here - #13226 (review)

Contributor

piyushghai commented Apr 5, 2019

@ChaiBapchya Are your concerns also addressed on this PR ?

Member

Roshrini commented Apr 17, 2019

@lobanov-m Can you please resolve conflicts?

lobanov-m added 3 commits

April 19, 2019 09:37


Added "factor" and "like" modes into BilinearResize2D operator.

f745f81

Also added tests and some fixes to visualization needed due to added modes.


Lint fix

01daea7


Test fix

52b2271

Contributor Author

lobanov-m commented Apr 19, 2019

@Roshrini rebase done.


retrigger CI

e95e535

wkcn removed the pr-awaiting-review label

wkcn approved these changes

View reviewed changes

Member

wkcn left a comment

All concerns have been addresses.
LGTM.
Thanks @lobanov-m for the excellent work!
I will merge it after the PR passes.

wkcn added 5 commits

April 28, 2019 22:07


retrigger CI

e24d907


Retrigger CI

37df076


retrigger CI

2d93b61


retrigger ci

6f55631


retrigger ci again

9b7b301

wkcn merged commit b30949f into apache:master

Member

wkcn commented May 5, 2019

Glad to merge it. Thank you!

arcadiaphy mentioned this pull request

flaky test: test_bilinear_resize_op #14902

Closed

Contributor Author

lobanov-m commented May 8, 2019

It's good news. Thank you!

access2rohit pushed a commit to access2rohit/incubator-mxnet that referenced this pull request


[MXNet-1211] Factor and "Like" modes in BilinearResize2D operator (ap…

29db90a

…ache#13226)

* Added "factor" and "like" modes into BilinearResize2D operator.
Also added tests and some fixes to visualization needed due to added modes.

* Lint fix

* Test fix

* retrigger CI

* retrigger CI

* Retrigger CI

* retrigger CI

* retrigger ci

* retrigger ci again

haohuanw pushed a commit to haohuanw/incubator-mxnet that referenced this pull request


[MXNet-1211] Factor and "Like" modes in BilinearResize2D operator (ap…

ad12334

…ache#13226)

* Added "factor" and "like" modes into BilinearResize2D operator.
Also added tests and some fixes to visualization needed due to added modes.

* Lint fix

* Test fix

* retrigger CI

* retrigger CI

* Retrigger CI

* retrigger CI

* retrigger ci

* retrigger ci again

lobanov-m deleted the bilinear-resize-scale-like branch

August 14, 2019 14:17

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.