Skip to content

Adds basic scripting support #81

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Merged
merged 18 commits into from
Jun 5, 2020
Merged

Adds basic scripting support #81

merged 18 commits into from
Jun 5, 2020

Conversation

narendasan
Copy link
Collaborator

@narendasan narendasan commented Jun 3, 2020 •
edited
Loading

Description

This PR will add support for scripted feed forward models like ResNet and MobileNet. It also changes how the evaluator system works, now you can set filters on evaluators to limit their scope which we use to separate loops we want to evaluate from loops we want to convert.

New support for:

  • prim::NumToTensor
  • aten::zeros
  • aten::mul.int
  • aten::sub.int
  • aten::__round_to_zero_floordiv
  • aten::slice.t
  • aten::len.t
  • prim::min.self_int

New lowering pass for:

  • Fusing addmm and matmul / add branches

Closes #16

Type of change

Please delete options that are not relevant and/or add your own.

  • New feature (non-breaking change which adds functionality)
  • This change requires a documentation update

Checklist:

  • My code follows the style guidelines of this project
  • I have performed a self-review of my own code
  • I have commented my code, particularly in hard-to-understand areas and hacks
  • I have made corresponding changes to the documentation and have regenerated the documentation (make html in docsrc)
  • I have added tests to verify my fix or my feature
  • New and existing unit tests pass locally with my changes

narendasan added 7 commits May 31, 2020 15:43
@narendasan
feat(//core/lowering): Fuse aten::addmm branches into a single
68f0317 
aten::addm op that can be expanded by a later pass

Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan
docs(//core/lowering): Document new aten::addmm fusion pass
17099fa 
Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan
feat(prim::NumToTensor): Implement evaluator for NumToTensor
60df888 
Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan
feat(aten::zeros): Implement aten::zeros evaluator
670817c 
Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan
feat(aten::to): Remove remaining typecast operators (should be a very
0f63ffa 
late pass to run)

Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan
docs: Update docs on new lowering passes
2f394fb 
Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan
feat(//core/conversion/evaluators): Adds new applicability filters for
2cc3226 
evaluators.

This allows developers to specifically blacklist or whiteline specific
cases of node kinds so that they will run on a subset of cases instead
of any instance. This is important in the case of prim::Loop where we
want to evaluate some loops and not others. This also lets us use
function schemas to target node, for instance there is now a
aten::mul.Tensor converter and an aten::mul.int evaluator.
In Tensor cases the converter will be called, in int cases the evaluator
will. We cannot switch to keying on function schema like we do for
converters because some node kinds dont have a schema so we do schema
white listing instead.

This commit also adds the following evaluators:

- aten::mul.int
- aten::sub.int
- aten::__round_to_zero_floordiv
- aten::slice.t
- aten::len.t
- prim::min.self_int

Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan narendasan marked this pull request as draft June 3, 2020 19:58
@github-actions github-actions bot added component: conversion Issues re: Conversion stage component: evaluators Issues re: Specific op evaluators component: lowering Issues re: The lowering / preprocessing passes labels Jun 3, 2020
narendasan added 2 commits June 3, 2020 13:55
@narendasan
fix(aten::size, other aten evaluators): Removes aten::size converter in
c83447e 
favor of an evaluator. Also fixes a bunch of bugs with the evaluators

Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan
fix(//core/lowering): Conv2D -> _convolution pass was triggering conv
ca2b5f9 
transpose instead of conv

Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@github-actions github-actions bot added the component: converters Issues re: Specific op converters label Jun 3, 2020
narendasan added 6 commits June 4, 2020 17:05
@narendasan
feat(//core/lowering): Adds peephole optimization pass
0014b84 
Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan
feat(//core/conversion/evaluators): adding support for common evaluation
d351717 
ops

Adds support for:
- prim::shape
- aten::neg
- aten::add
- aten::__getitem__
- aten::append

Fixes:
- prim::min

Removes:
- prim::Loop

Signed-off-by: Naren Dasan <[email protected]>
@narendasan
feat(//core/conversion): Adds the ability to evaluate loops
dcb1474 
Note: This is ungaurded currently, loops either must be able to be
evaluated at compile time or the module is not supported. Upcomming work
on RNNs will add support for more types of loops

Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan
fix(//core): Do not compile hidden methods
6bd1a3f 
(methods prefixed with _)

Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan
tests: Add scripted modules to the testsuite
3abfcaf 
Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan
docs: Update docs on new lowering passes
961f9d8 
Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan narendasan marked this pull request as ready for review June 5, 2020 00:23
narendasan added 3 commits June 4, 2020 17:36
@narendasan
refactor(//core): Update some typos and small code changes for the PR
5b2a2ec 
Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan
refactor(//core/lowering): Uncomment RemoveTo
111857c 
Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan
docs: One pass left out, links were wrong
80d5069 
Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]>
@narendasan narendasan merged commit 6d60246 into master Jun 5, 2020
@narendasan narendasan deleted the fuse_addmm_branches branch June 5, 2020 00:43
frank-wei pushed a commit that referenced this pull request Jun 4, 2022
Fix the FUSE_PASSES_POST_OBSERVER call (#81)
4a0ca3e 
Summary:
Pull Request resolved: https://github.com/pytorch/fx2trt/pull/81

The observer.observe calling signature is observe(module, inputs). But doing partial(observe, inputs) will reverse the order to observe(inputs, module).

Fixing it by using an explicit lambda.

Reviewed By: wushirong

Differential Revision: D36638291

fbshipit-source-id: 64abc0952802d5b438e1013c5ff91a57442900d9
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers
No reviews
Assignees
No one assigned
Labels
component: conversion Issues re: Conversion stage component: converters Issues re: Specific op converters component: evaluators Issues re: Specific op evaluators component: lowering Issues re: The lowering / preprocessing passes
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.

Start to handle branching in simple cases
1 participant
@narendasan