Support Caffe2 export/pure C(++) inference mode

[vivanov879]

Hi. Please suggest how to:

1. train a model in pytorch, save it on disk.
2. load the model in c++, call model:forward() several times without loading the model each time

[fmassa]

This thread might be relevant to you
https://discuss.pytorch.org/t/serving-model-trained-with-pytorch/168

[apaszke]

If I were to suggest something, I think I'd convert all the weights to numpy and dump them in a well established format like HDF5. These should be easily readable from C.

Our main idea is to add a model exporter, that could dump it to a Caffe2 graph, as that framework is very well optimized for production usage. And even though we don't plan to put a great amount of work on packaging production-ready models trained in PyTorch right now, there are some projects in the community that are aiming to do that, and we're going to provide them with support. You can find them in the forum link that @fmassa posted.

[taion]

Has the status or priority of this issue changed at all with the formal launch of Caffe2?

[soumith]

we're working on this, but it's still medium priority.

[ryanleary]

Any way to track progress on this or to collaborate in such a way that won't duplicate efforts @soumith?

[hasB4K]

Does that mean there is a way to get static graph information from PyTorch? Having an exporter to caffe2 would be great!

[jekbradbury]

This is in progress here #1815

[soumith]

fixed via http://pytorch.org/docs/master/onnx.html