Deep learning

Current implementation of Go binding can not specify options.

GPUOptions struct is in internal package. And go generate doesn't work for protobuf directory. So we can't specify GPUOptions for NewSession.

🚀 The feature, motivation and pitch

JIT serialization & freezing uses equality checks for things like common subexpression elimination (CSE).

Since aten::equal isn't implemented for sparse tensors, sparseTensorA.equals(sparseTensorB) fails, preventing CSE during freezing.

repro:

import torch

x = torch.rand((3, 4)).to_sparse()
y = torch.rand((3, 4))
print(torch.equal(x, 

Problem

When checkpointing a Torch model's state_dict, there may be some inconsistencies when saving/loading depending on whether the model is wrapped in DDP.

Proposal

Provide a utility method that always fetches the non-DDP version of the state_dict.

Without DDP:

model.state_dict()

With DDP:

model.module.state_dict()

Also see

[torch.nn.modu