torch backend LSTMRegressor fit fails

[simpsus]

File "E:\github\crowdcent\model.py", line 112, in objective model.fit( File "C:\Users\grumb\miniconda3\envs\crowdcent\Lib\site-packages\centimators\model_estimators\keras_estimators\sequence.py", line 47, in fit super().fit( File "C:\Users\grumb\miniconda3\envs\crowdcent\Lib\site-packages\centimators\model_estimators\keras_estimators\base.py", line 85, in fit X_np = _ensure_numpy(X) ^^^^^^^^^^^^^^^^ File "C:\Users\grumb\miniconda3\envs\crowdcent\Lib\site-packages\centimators\narwhals_utils.py", line 27, in _ensure_numpy return numpy.asarray(data) ^^^^^^^^^^^^^^^^^^^ File "C:\Users\grumb\miniconda3\envs\crowdcent\Lib\site-packages\torch\_tensor.py", line 1211, in __array__ return self.numpy() ^^^^^^^^^^^^ TypeError: can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.

What puzzles me about this is: It does not matter what I pass into fit. I try to pass a numpy directly to circumvent the narwhals fallback in _utils line 27. But somewhere ... no idea where, torch converts this into a GPU tensor. Any Idea where I could look?

[simpsus]

I set the backend with the import method. If I do not do that, it runs. Albeit with CPU, so no GPU, but it runs ...

[jrosenfeld13]

I see. We can tweak _ensure_numpy to conditionally handle torch tensors on GPU (if backend is torch do we check, or maybe there is a more general way) and transfer over to cpu as the error suggests

[jrosenfeld13]

It turns out there was a slightly more general way to fix this by switching to numpy ops instead of keras ops. Fixed in #10