[Relay, Torch] Clean up and refactor PyTorch frontend #4944
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
masahi
force-pushed
the
torch-refactor
branch
from
3063bc0
to
b8d334c
Compare
|
Thanks for doing this and looking forward to seeing the coming, additional PyTorch support :) |
masahi
force-pushed
the
torch-refactor
branch
from
fa7f6c3
to
10f4f9a
Compare
|
@alexwong The input name issue is fixed, see the last commit |
anijain2305
reviewed
Feb 26, 2020
masahi
force-pushed
the
torch-refactor
branch
2 times, most recently
from
d216c0b
to
98675af
Compare
|
Thanks for the changes. I will take a deeper look tomorrow. |
zhiics
reviewed
Feb 27, 2020
zhiics
reviewed
Feb 27, 2020
anijain2305
approved these changes
Feb 27, 2020
masahi
force-pushed
the
torch-refactor
branch
from
98675af
to
3a2cc94
Compare
|
Merging now, since my next PRs are pending. Thanks @anijain2305 @zhiics |
trevor-m
pushed a commit
to trevor-m/tvm
that referenced
this pull request
Apr 16, 2020
* The initial import of refactored implementation, all tests passed * enable mobilenet v2 test * minor cleanup * reorg * fix lint * use input names that come with torch IR * fix typo * introduce parse_operators * fix lint * add _ prefix
zhiics
pushed a commit
to neo-ai/tvm
that referenced
this pull request
Apr 17, 2020
* The initial import of refactored implementation, all tests passed * enable mobilenet v2 test * minor cleanup * reorg * fix lint * use input names that come with torch IR * fix typo * introduce parse_operators * fix lint * add _ prefix
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This is a follow up to #4497. The main motivation is to make it easier to add support for control flow and quantized ops later. This PR itself doesn't contain any new functionality. The diff for the upcoming control flow PR, building on top of the refactoring made in this PR, is here.
Remove the class boilerplate. There is no need for using a big class with many states in the frontend really. Instead, I made everything just a function. This departs from other frontends, but I think it is better suited for parsing Torch IR in particular. In my upcoming control flow PR, I need to use recursion to parse conditionals and loops, and it is naturally expressed in a functional style.
Simplify the parsing "main loop" by removing as much unnecessary intermediate variables as possible. Inputs to each relay op, constants and input types are retrieved on the fly. We now keep updating only two variables, the outputs of intermediate relay ops and indices into that outputs.
Simplify the
prim::GetAttrparsing logic inparse_params(...). The current implementation is difficult to understand and if you only look at the code it is not clear what it does.prim::GetAttris inherently recursive, so I use some recursion stuff to handle it. I think it is much clearer this way. As a bonus,get_use_chains(...)function I added can be used for other purpose as well.Please review @zhiics @icemelon9 @alexwong @jwfromm.
Aside from
prim::GetAttrmechanics, it doesn't require a detailed understanding of Torch IR.