[Relay][Frontend][ONNX] Fix reshape precompute, and type error #3230
Conversation
|
Please fix the lint error |
|
@jroesch can you look into the CI error? |
|
I disabled a few tests that require debugging version issues on CI. I propose we merge this, and I'll do a follow up pass afterwards. See my recent issue #3247. |
|
The CI dependency version issue is really a headache now. |
|
Yeah this is annoying, @zhreshold @tqchen can we bump the ONNX version? |
|
Ideally, we should make the code compatible with both versions. Please try to do that and let us merge it in. Please send a PR to https://github.com/dmlc/tvm/blob/master/docker/install/ubuntu_install_onnx.sh#L24 to ping the onnx to a specific version. then we can try to upgrade the image |
The problem is that the operator support in ONNX and in the torch export functionality varies across versions, so tests have to be conditionally enabled if we want to test model support. I believe it is important we test the entire pipeline on our CI because these models seem to have a lot of failures on the discuss board lately. |
|
#3374 updates the docker image to the latest state |
| check_torch_conversion(torchvision.models.resnet18, (1,3,224,224)) | ||
| # check_torch_conversion(torchvision.models.resnet101, (1,3,224,224)) | ||
|
|
||
| # def test_alexnet(): |
There was a problem hiding this comment.
What's the reason for these tests being commented out? Can we add a TODO if it's pending on some feature that isn't implemented yet?
|
Thanks @jroesch @zhreshold this PR has been merged. I need to fix issues in #3374 and will report back when that is ready |
|
Note: I reverted this PR for now due to CI errors, please open a new PR to bring things back, sorry about the trouble. |
|
The problem is because fusion was not performed here: |
|
NOTE, now that #3374 has been merged, you should be able to use the latest env |
apache#3230)" (apache#3385) This reverts commit df6957a.
apache#3230)" (apache#3385) This reverts commit df6957a.
Fix the code which precomputes shapes, and repair type error by casting literal integers to int32.