Extending @uni19's work on support dynamic shape input and scale_factor in interpolate layer #295
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
Signed-off-by: uni19 <[email protected]>
…into uni19-dynamic_interpolate
suite This commit adds onto the work done by @uni19 and adds support for the two variants of upsample ops. It also radically simplifies the tests for this component. Known issues: Right now both for scales and scale factors align corners being true causes errors Signed-off-by: Naren Dasan <[email protected]> Signed-off-by: Naren Dasan <[email protected]>
… edge case This commit adds support in some cases for the edge case when handling torch.nn.functional.interpolate where the user is doing some form of linear upsampling and uses scale factor to calculate the new tensor size at runtime and they set align corners to true (as of PyTorch 1.5 this is no longer the default behavior). This commit adds support for this case when users chose to construct static input size engines via the interpolate plugin which will run the function from ATen on CPU. In the case of dynamic input shapes with these 3 conditions the compilation will terminate with an error. The ultimate solution will be to find the root cause of the descripancy between PyTorch and TensorRT. Barring that we will need to use the dimension calculation primatives for TensorRT plugins. However, there is a limitation where static values in the computation cannot be floats which PyTorch scale factors are. Therefore it doesn't seem possible currently to support this usecase. Signed-off-by: Naren Dasan <[email protected]> Signed-off-by: Naren Dasan <[email protected]>
Signed-off-by: Naren Dasan <[email protected]> Signed-off-by: Naren Dasan <[email protected]>
github-actions
bot
added
component: conversion
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
peri044
approved these changes
Jan 28, 2021
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.
Description
This PR includes the changes in #258, does some refactoring of the test suite for interpolate and should expand coverage of the interpolate converter for the two variants of each of the currently supported resize functions (self, and vec). It should mean that we support interpolate for nearest and linear variants for all cases except when users use a type of linear upsampling, use scale factor for setting the size and set
align_cornersto true. It seems like @uni19 found that TensorRT and PyTorch have different behavior in this case. I will check with the TRT team to see if they have observed this. Hopefully it is not a massive issue as PyTorch since 1.5 has switched toalign_cornersbeing false by default. As a stop gap this PR extends the interpolate plugin to handle this case, however it can only do so for static TRT engines. Currently the build time dimension calculation can only be done withints where as scale factors for PyTorch arefloats. Therefore in the case of dynamic engine for a model using [Bi/Tri]Linear interpolate with scale factor andalign_corners=True, TRTorch will error out. Hopefully this is niche enough of a use case to not cause issues until we get a response from TensorRT.Fixes #290
Type of change
Please delete options that are not relevant and/or add your own.
Checklist: