Adding a Resnet 18 Example #268
Conversation
|
@IanNod or @aviator19941 can one of you TAL at this? |
| @@ -0,0 +1,3 @@ | |||
| transformers | |||
There was a problem hiding this comment.
Maybe worth considering moving resnet from examples to python/turbine_models in some form to enable use outside turbine?
examples/resnet-18/resnet-18.py
Outdated
| x = torch.randn(17, 3, 224, 224) | ||
| x[2] = pixel_tensor | ||
| y = shark_infer(x) | ||
| print_labels(y.to_host()) |
There was a problem hiding this comment.
Should add a test for this to ensure we don't break our example models
There was a problem hiding this comment.
I added something like a test, but I'm not sure if it is what you were requesting. I downloaded a bigger dataset and compared the results of the compiled module with the non SHARK-Turbine forward function.
There was a problem hiding this comment.
I was more thinking unit tests that will run on the CI so anytime our code is updated it will be checked for any kind of failures on this model. An example of the aot mlp example can be found here for reference https://github.com/nod-ai/SHARK-Turbine/blob/main/tests/examples/aot_mlp_test.py
There was a problem hiding this comment.
Amusingly enough, this example broke today upon updating the pip release of shark_turbine from 9.2 to 9.3.
There was a problem hiding this comment.
That is unfortunate, my condolences. This does highlight the fact that the tests are important as whatever change broke this example would have failed and would need to be addressed to get merged.
It may be helpful if you can post an issue of what the error is and maybe a minimal reproducer and we can get more eyes on the problem
There was a problem hiding this comment.
Will do! Thanks for all the comments so far, @IanNod
There was a problem hiding this comment.
Will do! Thanks for all the comments so far, @IanNod
Of course, thanks for the work you are doing!
|
@zjgarvey can you mark your test as xfail? we can merge this for now if you do |
This pull request adds a simple resnet-18 example with dynamic batches.