Add a Constant Propagation Pass to the JIT by eellison · Pull Request #8808 · pytorch/pytorch

eellison · 2018-06-22T22:09:55Z

Adding a constant propagation pass to the JIT. I have added examples to the expect files.

There are a couple of special cases which have not been implemented here. IF nodes with constant conditions can be inlined with the correct block. WHILE nodes can be removed if the condition is false. I have added a test for each case in test_jit.py file as expected failures.

To be consistent with DCE, python ops & CPP ops are treated as not having side-effects.

yf225 · 2018-06-22T22:17:13Z

There are probably files that we don't want to put in the PR. Can we rebase onto master, run git submodule update, and then push again?

apaszke

Mostly LGTM, but has some accidental checkins that will need to be removed

torch/csrc/jit/passes/constant_propagation.cpp

eellison · 2018-06-25T23:01:44Z

@apaszke Thanks for your comments! I will resubmit with the changes requested

eellison · 2018-06-27T15:24:51Z

@apaszke all the requested changes have been made.

@jamesr66a I ran git submodule init & force pushed, but it looks like third_party/onnx is still in the pull request.

ps: sorry about the commit clutter 😬 First time using git in a while, next pull request will be cleaner.

zdevito · 2018-06-29T17:23:53Z

This looks good, but we need to remove the changes to thirdy_party/onnx in the diff.
You want to do:

git diff origin/master third_party/onnx
cd third_party/onnx
git checkout ${onnx_commit_hash_of_master}
cd ../..
git commit -a --amend

soumith · 2018-07-09T22:11:19Z

in Convnet models, it's common for the final layer to have a view(batch_size, -1), like this resnet.

This shows up in the trace like this:

 %1871 : Float(4, 2048, 1, 1) = aten::avg_pool2d[kernel_size=[7, 7], stride=[1, 1], padding=[0, 0], ceil_mode=0, count_include_pad=1](%1869)
  %1872 : Long() = aten::size[dim=0](%1871)
  %1873 : Long() = prim::Constant[value={-1}]()
  %1874 : Dynamic = aten::stack[dim=0](%1872, %1873)
  %1875 : Float(4, 2048) = aten::view(%1871, %1874)
  %1876 : Float(2048!, 1000!) = aten::t(%319)
  %1877 : Float(4!, 1000) = aten::expand[size=[4, 1000], implicit=1](%320)

Since traces are already specialized on sizes (i.e. we know the size of %1871), it'll be good to fold these instructions in the Constant Folding pass.

There's a "just enough done" pass on doing this by @asuhan at https://github.com/asuhan/pytorch/blob/master/torch/csrc/jit/passes/constant_folding.cpp

yf225 · 2018-07-09T23:29:00Z

@eellison Could you rebase this onto current master? It should fix the No submodule mapping found in .gitmodules for path 'third_party/tbb' error.

zdevito · 2018-07-13T00:01:48Z

@eellison can we get this merged? It is going to start blocking other PR very soon.

eellison · 2018-07-25T00:44:10Z

@pytorchbot retest this please

eellison · 2018-07-25T01:22:58Z

@pytorchbot retest this please

facebook-github-bot

@eellison has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

ezyang · 2018-07-25T14:15:16Z

@pytorchbot retest this please

eellison · 2018-07-25T16:00:31Z

@pytorchbot retest this please

eellison · 2018-07-25T16:42:49Z

@pytorchbot retest this please

ezyang · 2018-07-25T17:14:20Z

@pytorchbot retest this please

ezyang · 2018-07-25T17:37:18Z

@pytorchbot retest this please

eellison · 2018-07-25T19:17:26Z

@pytorchbot retest this please

eellison · 2018-07-25T23:05:46Z

@pytorchbot retest this please

ezyang · 2018-07-26T17:26:45Z

this error looks real:

02:13:09 ======================================================================
02:13:09 ERROR: test_type_annotations_varargs (__main__.TestScript)
02:13:09 ----------------------------------------------------------------------
02:13:09 Traceback (most recent call last):
02:13:09   File "test_jit.py", line 2941, in test_type_annotations_varargs
02:13:09     self.checkScript(fn1, (x, y, z), optimize=True)
02:13:09   File "test_jit.py", line 1417, in checkScript
02:13:09     self.checkScript(source, inputs, optimize, outputs, script.__name__, capture_output, frames_up=2)
02:13:09   File "test_jit.py", line 1427, in checkScript
02:13:09     outputs_ge = ge(*inputs)
02:13:09 RuntimeError: 
02:13:09 Schema not found for node. File a bug report.
02:13:09 Node: %3 : Dynamic = ^fn_varargs()(%4)
02:13:09 
02:13:09 Input types:Dynamic
02:13:09 candidates were:
02:13:09 :
02:13:09 def fn1(x, y, z):
02:13:09     return fn_varargs(x)
02:13:09            ~~~~~~~~~~ <--- HERE
02:13:09 
02:13:09

eellison · 2018-07-27T17:32:45Z

@pytorchbot retest this please

eellison · 2018-07-27T19:15:06Z

@pytorchbot retest this please

eellison · 2018-07-30T16:39:56Z

@pytorchbot retest this please

zdevito · 2018-07-30T17:50:02Z

This is failing because it needs to be rebased and its expect tests need to be --accepted. The removal of attributes has changed how some graphs print.

facebook-github-bot

eellison has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

Summary: Adding a constant propagation pass to the JIT. I have added examples to the expect files. There are a couple of special cases which have not been implemented here. IF nodes with constant conditions can be inlined with the correct block. WHILE nodes can be removed if the condition is false. I have added a test for each case in test_jit.py file as expected failures. To be consistent with DCE, python ops & CPP ops are treated as not having side-effects. Pull Request resolved: pytorch#8808 Reviewed By: wanchaol Differential Revision: D8906770 Pulled By: eellison fbshipit-source-id: 10ad796d89f80b843566c9ddad6a0abd1f3dc74c

eellison requested review from apaszke, colesbury, ezyang, gchanan, orionr, soumith and zdevito as code owners June 22, 2018 22:09

apaszke suggested changes Jun 23, 2018

View reviewed changes

eellison force-pushed the constant_prop branch from b4f5f26 to 98d41f3 Compare July 9, 2018 20:50

eellison force-pushed the constant_prop branch from 57b6734 to 43d365d Compare July 9, 2018 22:54

eellison requested review from Yangqing, anderspapitto, bddppq, dzhulgakov, houseroad, jamesr66a, pietern, smessmer and teng-li as code owners July 10, 2018 16:37

eellison closed this Jul 10, 2018

zdevito reopened this Jul 13, 2018

eellison force-pushed the constant_prop branch 2 times, most recently from 683c3fd to 20b30ff Compare July 17, 2018 16:58

eellison force-pushed the constant_prop branch 2 times, most recently from 64fd978 to 9e5fc58 Compare July 25, 2018 01:51

facebook-github-bot reviewed Jul 25, 2018

View reviewed changes

eellison force-pushed the constant_prop branch 3 times, most recently from 583b149 to f2dbfd5 Compare July 27, 2018 03:50

eellison force-pushed the constant_prop branch from f2dbfd5 to 59501a4 Compare July 27, 2018 19:06

eellison force-pushed the constant_prop branch from 59501a4 to 45a3d2d Compare July 27, 2018 19:46

Elias Ellison added 2 commits July 30, 2018 11:50

Constant prop commit, cleanedup.

32ee903

updating for attributes no longer being used

2f5d900

eellison force-pushed the constant_prop branch from 45a3d2d to 2f5d900 Compare July 30, 2018 19:12

facebook-github-bot reviewed Jul 30, 2018

View reviewed changes

facebook-github-bot closed this in e57cb4a Jul 30, 2018

ezyang added the merged label Jun 26, 2019

Conversation

eellison commented Jun 22, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

yf225 commented Jun 22, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

apaszke left a comment

Choose a reason for hiding this comment

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!

eellison commented Jun 25, 2018

Uh oh!

eellison commented Jun 27, 2018

Uh oh!

zdevito commented Jun 29, 2018

Uh oh!

soumith commented Jul 9, 2018

Uh oh!

yf225 commented Jul 9, 2018

Uh oh!

zdevito commented Jul 13, 2018

Uh oh!

eellison commented Jul 25, 2018

Uh oh!

eellison commented Jul 25, 2018

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

ezyang commented Jul 25, 2018

Uh oh!

eellison commented Jul 25, 2018

Uh oh!

eellison commented Jul 25, 2018

Uh oh!

ezyang commented Jul 25, 2018

Uh oh!

ezyang commented Jul 25, 2018

Uh oh!

eellison commented Jul 25, 2018

Uh oh!

eellison commented Jul 25, 2018

Uh oh!

ezyang commented Jul 26, 2018

Uh oh!

eellison commented Jul 27, 2018

Uh oh!

eellison commented Jul 27, 2018

Uh oh!

eellison commented Jul 30, 2018

Uh oh!

zdevito commented Jul 30, 2018

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants

eellison commented Jun 22, 2018 •

edited

Loading

yf225 commented Jun 22, 2018 •

edited

Loading