Can I load Torch trained model file by Pytorch

[erogol]

I want to use torch trained deep learning model from python. Is it able to do that?

[hughperkins]

Yes ... you can create a lua class, that trains your model, or loads an existing model, and call this from Python. You can pass in new examples from python to the model, to predict on. I would first have a look at simpleexample/pybit.py and simpleexample/luabit.lua for an example of python code that loads a lua class, creates an object from it, and calls methods on it. That lua class could load your model, and pass new examples to your model for prediction.

[hughperkins]

Actually, I will write an example for this :-)

[erogol]

Thanks :)

[hughperkins]

Ok, I didnt add prediction, but covers:

• building a network
• passing in training data, in batches
• training the network
• getting training loss and accuracy
• also handles: cpu, cuda, opencl :-)

51cbc09

https://github.com/hughperkins/pytorch/tree/master/examples/luamodel

[hughperkins]

Updated so you can pass numpy tensors in directly now 66bb797

[hughperkins]

added prediction in 9c05864

Also added methods tensorOne.eq(tensorTwo), and tensor.sum() to PyTorch.xxxTensor objects

You'll need to pip install docopt to use runner.py now

[hughperkins]

Updated in 0a9e04a to convert the output predictions tensor to numpy array, before calculating number correct

[hughperkins]

To what extent does this do approximately what you want? (The other approach is to instantiate nn modules directly from the python; depends what you want really)

[hughperkins]

You can load directly now, without needing to create a lua class, see https://github.com/hughperkins/pytorch/issues/15