ArrayDataAdapter will needlessly convert backend native tensors

[mattdangerw]

Given the following snippet on the tf backend....

x = tf.ones((1000, 2))
y = tf.ones((1000,))
model.fit(x, y, batch_size=8)

The ArrayDataAdapter will convert all tf.Tensor inputs to np.array, before converting them back to the tf.Tensor format. The same roundtrip through np.array is true for jax and torch tensors on the respective backends.

This seem inefficient. We should benchmark if this is causing a slowdown on any backend, and fix if necessary.

[mattdangerw]

@haifeng-jin cc'ing you in case you have already thought on this.

This worst case seems like torch potentially with explicit device placement. We might be getting as input torch tensor on a CUDA device, converting them to numpy, and then converting back to tensors on the CUDA device. This seems like it might be slow (without any data).

[hertschuh]

@mattdangerw ,

I noticed this too. It seems like they could be converted only if needed and just in time as long as the tensor/array is sliceable. I'll be working on that class for other reasons so I'll take a look at that point.

[haifeng-jin]

I will work on this as I resolve PyTorch backend memory issues.