kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-693093647
@masahi Yeah. Those ops look like coming from scripted model. I believe for
pt 1.6 if we tracing the model there are 2 or 3 ops missing.
-
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-693016775
@masahi Yeah. It probably worth a forum discussion thread before we proceed
with any solution. In tf frontend, there are even more ```try .. except```
blocks which doe
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-693007551
@t-vi For now IMO it doesn't bring obvious benefit of adding an extra common
api to just wrap ```try ... except``` for ```_infer_value```, considering the
lack of gene
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-692999233
> But which part raises the exception? infer_value_ or the casting that
follows?
It's for ```_infer_value```.
-
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-692998861
@masahi I rolled back topk to still _infer_value. Directly use dyn topk will
cause a certain shape dim to be dynamic and cause get_valid_count issue.
Compiling pt mask
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-692985102
As I have discussed with @masahi, the problem of having a try interface is
that there is no common logic between different dynamic ops when dealing with
dynamic attrib
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-692975419
Original pt frontend just handles limited cases, mostly static
shape/attributes. It is fine we just keep input as it is. For more dynamic
models, we need to do some ex
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-692921717
@masahi It looks like certain recent change causes this error. I'm
investigating.
This is an automate
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-692917238
> @zhiics @kevinthesun @masahi
> Thank you, @kevinthesun for your summary and all the work in the
investigation and your PR.
>
> I think using `if isinstance(
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-692392769
> Is moving to fully dynamic frontend like being done for onnx #6351 help
removing infer_value usage?
While we try to make relay expression as static as possible
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-692373767
@zhiics @masahi @t-vi Sure. One major thing in this PR is the handling for
dynamic operators such as slice, arange and topk. These ops have dynamic
attribute which aff
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-692365187
E2E tests added. Now waiting for
https://github.com/apache/incubator-tvm/pull/6464.
This is an automa
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-692285615
> But we should still know when it is appropriate to run the inference,
should we not?
> I must admit I'm still troubled by the idea that we don't know beforehand
w
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-692260607
@masahi The problem of creating a try_infer_value API is that it doesn't
simplify the codes since we need to do various handling in except block for
different ops. We
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-690898948
This is an automated message from the Apache Git Service.
To respond to the message, please log on to G
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-690898948
This is an automated message from the Apache Git Service.
To respond to the message, please log on to G
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-691321674
@masahi Coming soon. :D
This is an automated message from the Apache Git Service.
To respond to the me
kevinthesun commented on pull request #6449:
URL: https://github.com/apache/incubator-tvm/pull/6449#issuecomment-690898948
@masahi These changes are mainly for torch vision rcnn models which enhances
current features. There is another PR adding some backend stuffs. After that I
can add e2e
18 matches
Mail list logo
