AdamW bug on v0.96.0

[QingtaoLi1]

When I update TorchSharp from 0.95.4 to 0.96.0, some exception occurs on torch.optim.AdamW():

System.Runtime.InteropServices.SEHException: 'External component has thrown an exception.'

The only change I commit is lower-casing train() and eval(). Notice that the first argument of torch.optim.AdamW() (and other optimizers) changes from IEnumerable<torch.Tensor> to IEnumerable<TorchSharp.Modules.Parameter>, so I guess the bug is related to library changes to solve #501.

[NiklasGustafsson]

@QingtaoLi1 -- the renaming and the type change should not affect anything. Down in native code, neither matters.

Could you provide the whole stack trace? The one line you have is not enough to give me an idea of what is going on.

[NiklasGustafsson]

Also, in the next release, AdamW will be implemented in managed code, so things are likely to change.

[QingtaoLi1]

@QingtaoLi1 -- the renaming and the type change should not affect anything. Down in native code, neither matters.

Could you provide the whole stack trace? The one line you have is not enough to give me an idea of what is going on.

Stack trace here. It points to the constructor of Adam and nothing else. The only error message is "external component..." I showed above.

[NiklasGustafsson]

Okay, thanks!

[QingtaoLi1]

@NiklasGustafsson Oh I didn't notice I was captioning the error of Adam. But the exception is the same for both optimizers.

[NiklasGustafsson]

Let's see how this changes with the new implementation in 0.96.1

[QingtaoLi1]

Let's see how this changes with the new implementation in 0.96.1

Great!

P.S. I tried SGD just now and it also fails.

[NiklasGustafsson]

Hmmm. There's something fundamentally wrong. Is there any possibility of any mixup of DLLs or packages?

[QingtaoLi1]

I'm not sure. Just try deleting .nuget\\packages\\torchsharp\\ and re-installing it, but nothing changes.

[NiklasGustafsson]

OK. Can you try running some of the examples in dotnet/TorchSharpExamples?

[QingtaoLi1]

@NiklasGustafsson Weird, TorchSharpExamples for AG_NEWS and MNIST can run. I don't find anything different between my code and the examples -- getting parameters by model.parameters() and right then building optimizers using torch.optim.SGD/Adam.

[NiklasGustafsson]

If you have a chance to narrow it down to a smaller example, running with random data, etc., that would be great. Maybe it has something to do with the specific modules you are using?