[MXNET-978] Add higher order gradient support `tan`, `tanh` by kshitij12345 · Pull Request #15253 · apache/mxnet

kshitij12345 · 2019-06-16T06:38:24Z

Description

Trying to add higher order support for tan, tanh.

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA-978 issue created (except PRs with tiny changes)
Changes are complete (i.e. I finished coding on this PR)
All changes have test coverage:
Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
Code is well-documented:
For user-facing API changes, API doc string has been updated.
For new C++ functions in header files, their functionalities and arguments are documented.
For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
Check the API doc at http://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

Support Higher order gradient for Tan and Tests
Support Higher order gradient for Tanh and Tests

…to add-higher-ord

kshitij12345 · 2019-06-16T06:47:36Z

@larroy @apeforest Please help.

I am facing some issue here.

https://github.com/apache/incubator-mxnet/blob/f49013a60b8ea5d6c75ba3515f25a8c346748269/src/operator/tensor/elemwise_unary_op_basic.cc#L123-L137

Here I am returning input[1] * ograd as value for x_grad_grad which I expect it to be f(x) * ograd as we use ElemGradUseOut to wrap the _backward_sigmoid function.

However from the tests.

https://github.com/apache/incubator-mxnet/blob/f49013a60b8ea5d6c75ba3515f25a8c346748269/tests/python/unittest/test_higher_order_grad.py#L109-L123

https://github.com/apache/incubator-mxnet/blob/f49013a60b8ea5d6c75ba3515f25a8c346748269/tests/python/unittest/test_higher_order_grad.py#L136-L151

You can see that the actual value returned is f(x) * f'(x) * ograd.
I have also tried the method as suggested by @larroy of using create_graph=False and retain_graph=True.

Thank You.

apeforest · 2019-06-17T23:47:26Z

@kshitij12345 I will review it carefully later tonight. Could you please also rebase your other PR #15120 and trigger CI? I need your update in the unit tests for my PR as well. Thanks!

apeforest · 2019-06-18T04:55:34Z

+
+    std::vector<nnvm::NodeEntry> ret;
+
+    ret.emplace_back(MakeNode("elemwise_mul", n->attrs.name + "_backward_grad_grad_inp",


Shouldn't the first item be _backward_grad_grad and the second item be _backward_grad_grad_inp?

Given that:

n->input[0]: ygrad n->input[1]: y

Actually the code is incorrect in terms of naming the attribute and probably variable names.
As my main concern was trying to figure why we f(x) * f'(x) where f(x) was expected.

kshitij12345 · 2019-06-18T05:02:22Z

@kshitij12345 I will review it carefully later tonight. Could you please also rebase your other PR #15120 and trigger CI? I need your update in the unit tests for my PR as well. Thanks!

Done.

apeforest · 2019-06-18T05:07:53Z

+
+    ret.emplace_back(MakeNode("elemwise_mul", n->attrs.name + "_backward_grad_grad_inp",
+                             {ograds[0], gx_ograd}, nullptr, &n));
+    ret.emplace_back(MakeNode("elemwise_mul", n->attrs.name + "_backward_grad_grad",


Shouldn't this be ograds[0] * f'(x) ?

Let f(x) = sigmoid(x)
xgrad = dL/dx
ygrad = dL/dy
We know dydx = y*(1-y)
In the first backward pass:

xgrad = dydx * ygrad n->inputs[0] ---> ygrad n->inputs[1] ---> y = f(x) nnvm::NodeEntry{n} ---> xgrad

According to the multiplier rule, in the second backward pass:

_backward_grad_grad = d2L/dx2 * f'(x) = ograds[0] * dydx _backward_grad_grad_inp = d2L/dx2 * ygrad * f''(x) = ograds[0] * n->input[0] * f''(x)

That is indeed the surprising part, I have asserted from Python that n->inputs[1] is f(x) * f'(x) surprisingly instead of it being f(x) that is why i was very much confused. And also if I remember correctly n->inputs[0] as well as nnvm::NodeEntry{n} are different from what we expect them to be.

Note : Right now it doesn't actually implement the second order gradient for sigmoid.

I am essentially returning what I expect to be f(x) * ograds(from second backward)
at line
https://github.com/apache/incubator-mxnet/blob/f49013a60b8ea5d6c75ba3515f25a8c346748269/src/operator/tensor/elemwise_unary_op_basic.cc#L134-L135
but instead we are receiving f(x) * f'(x) * ograds(from second backward) which is confirmed from Python. That is where my confusion lies.

I got your question now. I got the same result f(x) * f'(x) * ograds locally. Need to dig in more...

I dumped out the computation graph. If you want to try, refer to my local branch: https://github.com/apeforest/incubator-mxnet/blob/develop/higher_order_grad/src/imperative/imperative.cc#L509

[16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:300: node 0 var [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:300: node 1 var [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:302: node 2 sigmoid [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:306: input 1: [3] (0 KB) -> [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:312: output 2: [3] (0 KB) -> [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:302: node 3 _backward_sigmoid [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:306: input 0: [3] (0 KB) -> [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:306: input 2: [3] (0 KB) -> [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:312: output 3: [3] (0 KB) -> [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:300: node 4 var [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:302: node 5 elemwise_mul [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:306: input 4: [3] (0 KB) -> [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:306: input 2: [3] (0 KB) -> [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:312: output 5: [3] (0 KB) -> [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:302: node 6 _backward_sigmoid [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:306: input 5: [3] (0 KB) -> [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:306: input 2: [3] (0 KB) -> [16:37:15] /Users/lnyuan/work/mxnet/src/executor/../common/exec_utils.h:312: output 6: [3] (0 KB) ->

So it seems if the backward operator takes output of the forward node instead of input, the backward graph may be created differently with some chain rule applied already.

Oh Thanks. Will try to look into it. Thanks for the pointer.

larroy · 2019-06-18T22:37:34Z

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

larroy · 2019-06-18T22:37:51Z

@mxnet-label-bot add [operator]

…to add-higher-ord

update code as per apache#15288.

…to add-higher-ord

larroy · 2019-07-10T19:48:32Z

+      // f''(x) = 2 * f'(x) * f(x)
+      const std::unordered_map<std::string, std::string> args = {{"scalar", "2.0"}};
+      auto two_y = MakeNode("_mul_scalar", n->attrs.name + "_mul_two", {n->inputs[1]}, &args, &n);
+      auto grad_grad_mid = MakeNode("elemwise_mul", n->attrs.name + "_grad_mul",


Can we clarify / add a comment on why is correct to multiply by y_grad (the first head gradient?) again? This would help readers as is not obvious, as well as the very non-obvious implicit multiplication by f'(x) it compounds.

Makes sense. Thanks. Will get to it.

I have updated the comment. See if it is okay? Or maybe the phrasing can be improved.
Thanks

Thanks. About the outputs, I think we should write some documentation explaining what we are doing as I find it non trivial. Can you help me understand the y_grad_grad (first output)?

If you want, we can move the conversation to the dev list or slack, as the PR LGTM.

Ok I clarified with @apeforest , this makes sense now.

larroy

LGTM but I find the code non-obvious, (is consistent with sigmoid) I think we should add some documentation on the faq about this mechanics, the implicit multiplication by f'(x) and the multiplication by y_grad.

larroy · 2019-07-11T19:32:57Z

+      // f''(x) = 2 * f'(x) * f(x)
+      const std::unordered_map<std::string, std::string> args = {{"scalar", "2.0"}};
+      auto two_y = MakeNode("_mul_scalar", n->attrs.name + "_mul_two", {n->inputs[1]}, &args, &n);
+      auto grad_grad_mid = MakeNode("elemwise_mul", n->attrs.name + "_grad_mul",


Thanks. About the outputs, I think we should write some documentation explaining what we are doing as I find it non trivial. Can you help me understand the y_grad_grad (first output)?

If you want, we can move the conversation to the dev list or slack, as the PR LGTM.

kshitij12345 · 2019-07-14T07:24:34Z

It is actually confusing.
Sure.
Thank You.

larroy

LGTM

sandeep-krishnamurthy · 2019-07-20T05:47:28Z

@apeforest - Can you please take look at this PR and merge if it looks good? Thanks

apeforest

LGTM

…5253) * init to reset * issue: higher order backward sigmoid * update gradient code. update code as per apache#15288. * undo changes * relax tolerance of gradient mismatch for tanh * update comments * update comments

kshitij12345 added 3 commits June 1, 2019 21:57

init to reset

08042cc

Merge branch 'master' of https://github.com/apache/incubator-mxnet in…

3e76599

…to add-higher-ord

issue: higher order backward sigmoid

f49013a

apeforest self-requested a review June 17, 2019 23:31

apeforest reviewed Jun 18, 2019

View reviewed changes

marcoabreu added Autograd pr-awaiting-review PR is waiting for code review labels Jun 18, 2019

marcoabreu added the Operator label Jun 18, 2019

kshitij12345 added 2 commits July 2, 2019 23:18

Merge branch 'master' of https://github.com/apache/incubator-mxnet in…

6168da0

…to add-higher-ord

update gradient code.

9bbeb16

update code as per apache#15288.

kshitij12345 changed the title ~~Add higher order gradient support sigmoid, tan, tanh~~ Add higher order gradient support tan, tanh Jul 2, 2019

kshitij12345 changed the title ~~Add higher order gradient support tan, tanh~~ [MXNET-978] Add higher order gradient support tan, tanh Jul 2, 2019

kshitij12345 added 3 commits July 3, 2019 00:54

undo changes

9bc96f3

relax tolerance of gradient mismatch for tanh

8708a63

Merge branch 'master' of https://github.com/apache/incubator-mxnet in…

f91de79

…to add-higher-ord

larroy reviewed Jul 10, 2019

View reviewed changes

larroy approved these changes Jul 10, 2019

View reviewed changes

update comments

ed1ace8

larroy approved these changes Jul 11, 2019

View reviewed changes

larroy approved these changes Jul 16, 2019

View reviewed changes

update comments

6f0bb50

apeforest approved these changes Jul 29, 2019

View reviewed changes

apeforest merged commit 8e31dad into apache:master Jul 29, 2019

kshitij12345 deleted the add-higher-ord branch July 30, 2019 14:04


		std::vector<nnvm::NodeEntry> ret;

		ret.emplace_back(MakeNode("elemwise_mul", n->attrs.name + "_backward_grad_grad_inp",

Conversation

kshitij12345 commented Jun 16, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Checklist

Essentials

Changes

Uh oh!

kshitij12345 commented Jun 16, 2019

Uh oh!

apeforest commented Jun 17, 2019

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kshitij12345 commented Jun 18, 2019

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kshitij12345 Jun 18, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

larroy commented Jun 18, 2019

Uh oh!

larroy commented Jun 18, 2019

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

larroy left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kshitij12345 commented Jul 14, 2019

Uh oh!

larroy left a comment

Choose a reason for hiding this comment

Uh oh!

sandeep-krishnamurthy commented Jul 20, 2019

Uh oh!

apeforest left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

kshitij12345 commented Jun 16, 2019 •

edited

Loading

kshitij12345 Jun 18, 2019 •

edited

Loading

larroy left a comment •

edited

Loading