Using experimental features

1

Hello PyTorch community,

I’d like to use experimental functions such as torch.vmap() which are not included in the stable version, how can I do so ?

Best,

Luis

2

You can install the Preview (nightly) binaries from here to get the latest build. :slight_smile:

3

Just installed pytorch-nightly in a dedicated conda env for GPU and then CPU only (since the first option didn’t work, I suspected a downgrade imposed by my cuda 10.1) in order to use torch.vmap but it seems I don’t have the right nightly: I keep generating the error

AttributeError: module ‘torch’ has no attribute ‘vmap’

Furthermore, a

import torch
print(torch.__version__)

in the notebook launched with the new nightly dedicated conda env returns 1.7.1, although the nightly should be 1.8.X according to this post. I tried installing pytorch nightly successively with the command lines suggested on the pytorch website homepage:
conda install pytorch torchvision torchaudio cudatoolkit=10.1 -c pytorch-nightly
conda install pytorch torchvision torchaudio cpuonly -c pytorch-nightly

Probably unrelated but would it be linked to the Remove torch.vmap here ? According to this github issue torch.vmap should be available in pytorch 1.7 anyway.

I’m calling torch.vmap in the following line of code, part of a class method:
torch.vmap(self.distance_riemann_to_c)(net_outputs)
torch.vmap thus vectorizes the application of the other method (same class) self.distance_riemann_to_c to the batched nnet outputs net_outputs. Sorry if it would have been more appropriate to open a new post (?).

4

If torch.__version__ returns 1.7.1, I guess you might have multiple PyTorch installations in the current environment and your kernel picks the 1.7.1 one.
Could you either uninstall other PyTorch installations or create a new environment and install the nightly there?

1 Like
5

This is what I tried. However, today when I re-created a CPU nightly env, and a GPU nightly env, I wasn’t able to reproduce my issue & got the expected 1.9.0dev, and have obtained access to torch.vmap(). I must have done something wrong the first time ! Thanks for the quick answer :slight_smile: