06 transfer learning using torch.compile()

[sacromentus]

I tried implementing torch.compile() on the EfficientNet model, i did model = torch.compile(model) right after I instantiate model, but it does not work. First, torchsummary doesn't work with it, which I read is normal. But when I train the model, it won't train and I get various errors, one of them being "BackendCompilerFailed: backend='inductor' raised:"

Has anyone implemented it successfully?

[mrdbourke]

Hi @sacromentus ,

I haven't tried too much transfer learning with torch.compile() just yet.

But I would like to.

Have you tried a different backend with torch.compile()? See: https://pytorch.org/docs/stable/generated/torch.compile.html

Also, which hardware are you running on? That might effect how torch.compile() performs?

Have you managed to get the model to do a single forward pass? That might help with troubleshooting.

[sacromentus]

Daniel,

Thanks for the reply and the amazing course! I will try a different backend; also, I will try it on some of the models we built instead of a pre-trained one. I’ll report back my results.My hardware is a laptop NVIDIA 4070 (cuda is enabled, and boy was that a learning experience in and of itself!)

-Jerry

[mrdbourke]

Hahaha you're so right! Setting up CUDA for the first time can be a bit of a challenge.

Glad to hear you got it working :)

Also I'd be keen to hear if you get torch.compile() working with a pre-trained model too!