[minor] Use torchvision for data ?

[dmudiger]

PyTorch-LBFGS/examples/full_batch_lbfgs_example.py

Line 42 in 6677e1f

(X_train, y_train), (X_test, y_test) = cifar10.load_data()

Use torchvision or something similar for data, instead of having to have keras (+TF/Theano backend) for just the data loading :)

[hjmshi]

OK, let me see if I can change this. The issue with using the Torchvision dataloader is that it is not easily compatible with multi-batch L-BFGS, unless we want to write a dataloader specifically for that. It should work though for full-batch and full-overlap L-BFGS. Is there a way to obtain the dataset in tensor format using torchvision? We could use this to load the dataset then do all of the data loading manually for the optimization algorithm. What do you think?

[dmudiger]

Full-batch should be quite straightforward, but for multi-batch it might need additional changes. In that case we can just stay with this, its not a big deal.

[hjmshi]

Let me see if I can still load the dataset though through torchvision (without using the dataloader).

[hjmshi]

It is surprisingly annoying to load the dataset as a tensor through torchvision...do you know how to do this? Also, do you know if PyTorch plans on releasing a dataloader with adaptive batch sizes? If they release something like that, that would make multi-batch much easier to implement (and progressive batching).