torchbearer

torchbearer copied to clipboard

Queston about training loop

4

Hi! I'm trying to fork the repo and add some functionality for an experiment. But that requires an addition in the training loop. I've read the documentation and the code...

AnabetsyR

 Please check the example of Learning Rate Schedulers.

SmallSquare

When I insert this callback in the trial I get the following error. Is this some kind of bug? It seems like the gradients are not passed in the callback....

dimimal

Native automatic mixed precision support (`torch.cuda.amp`) is now in master: https://pytorch.org/docs/master/amp.html https://pytorch.org/docs/master/notes/amp_examples.html Not sure if you ever tried Nvidia's (our) experimental Apex Amp, but I know it has many pain...

mcarilli

When y_pred is a tuple (i.e. model returns multiple outputs) the criterion now recieves the tuple unpacked. This should either be reverted or more clearly documented.

ethanwharris

Should include most of the experimented modes https://arxiv.org/abs/1805.11272

ethanwharris

Test JIT support and add an example

ethanwharris

For most callbacks we don't need to have a string representation, but for callbacks that modify the loss or callbacks that save files, it would be good to have one...

MattPainter01

If you pass a 4D tensor (rows, columns, height, width) to TensorBoardImages then it would be good if it could detect this and infer the make_grid parameters: num_images, nrow. Would...

MattPainter01

Would be nice if we could kill torchbearer training with a ping or similar

ethanwharris