Skip to content
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

Add operators necessary for mobilenet #29

Merged
merged 15 commits into from
Mar 31, 2020
Merged

Add operators necessary for mobilenet #29

merged 15 commits into from
Mar 31, 2020

Conversation

narendasan
Copy link
Collaborator

  • Adds reduce ops and hardtanh.
  • Enhances Args to check types before unwrapping.
  • Adds a method to the conversion context to reduce boiler plate in the converters
  • Conversion now checks to see that all Values have corresponding ITensors, and emits a warning otherwise. This means the contract for converters has changed, now they must declare an association for all output tensors for suppress this warning (even if its a nullptr).
  • ConvertersLUT key changed to c10::OperatorName to support operator overloading in TorchScript
  • Fixes a bug in linear which would include the batch size when flattening
  • Adds a slightly reduced threshold to module level tests

@narendasan
feat(//core/conversion/converters/impl/reduce): Mean reduce converter
259aa4c 
Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan
refactor(//core/conversion/converters): Removing a bunch of boilerplate
8ad8fa7 
and add some checking

Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan
feat(//core/conversion/conversionctx): Move inline function to associate
6ab9814 
values and tensors to a method in the conversion ctx

Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan
refactor(//core/conversion/converters): Move to key converters by
298b5bb 
c10::OperatorName allowing support for different converters for
overloaded operators

Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan
refactor(//core/conversion): Now check to make sure all node outputs
de8659b 
have tensor matches, and move to checking with execptions instead of
bool returns for the conversion stage

Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan
feat(//core/conversion/converters/impl): Non dimensional reduce
ccab7b9 
converter

Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan
doc(//tests): Add suggested settings for running the full test suite so
36b2eb4 
as to not exhaust GPU memory

Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan
fix(//core/conversion/converters/impl/linear): In inserting flatten for
377ad67 
layers other than conv, did not remove batch size from tensor shape

Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan
fix(//core/conversion/converters/impl/reduce): Adds support for multiple
7622a97 
reduction dimensions

Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan
refactor(//tests): Relax threshold for module tests
d1270d9 
Perhaps revisit in the future to see if we can reduce this back

Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan
fix(//tests/modules): Remove an old script
8be79e1 
Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan
feat(//core/conversion/converter/Arg): Add typechecking to the unwrap
73bfd4c 
functions

Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan
feat(hardtanh): Adds support for the the hard tanh operator
391af52 
Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan
feat(//core/conversion/converters/impl/reduce): adds the rest of TRT's
956b0c5 
basic reduction operators

Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan narendasan added component: tests Issues re: Tests component: core Issues re: The core compiler component: converters Issues re: Specific op converters labels Mar 31, 2020
@narendasan
refactor(//core/conversion/converters/impl/element_wise): remove some
ca42ef5 
old comments

Signed-off-by: Naren Dasan <naren@narendasan.com>
Signed-off-by: Naren Dasan <narens@nvidia.com>
@narendasan narendasan merged commit aef6003 into pytorch:master Mar 31, 2020
@narendasan narendasan deleted the mobilenet branch March 31, 2020 07:44
@narendasan narendasan linked an issue Apr 1, 2020 that may be closed by this pull request
Solidify Arg class #3
Closed
2 tasks
frank-wei pushed a commit that referenced this pull request Jun 4, 2022
[fx2trt] support masked_fill, repeat (#29)
eee660d 
Summary:
Pull Request resolved: https://github.com/pytorch/fx2trt/pull/29

BERT_pytorch model in torchbench:
1. masked_fill, repeat
2. fix a nit in embedding op

Reviewed By: yinghai, wushirong

Differential Revision: D35034664

fbshipit-source-id: 2c0e66df5a17f6960d71c3bc6e9bdca9a5daf3e5
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: converters Issues re: Specific op converters component: core Issues re: The core compiler component: tests Issues re: Tests
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.

Solidify Arg class
1 participant
@narendasan