Is it possible to load pytorch model (.pth)?

[theheena]

Hi,
I have a pytorch pre-trained model (.pth) for neural art style. Is it possible to convert this model into mlmodel?

[opedge]

Hi @theheena ,
Although torch7 and pytorch seem similar, models are not compatible with each other. Only torch7 -> pytorch (legacy.nn) is suitable for conversion. But not vice versa. Thus it is not possible to convert .pth pre-trained pytorch model.

[theheena]

Hi @opedge ,

Thank you for your answer.

[theheena]

Hi @opedge ,

I have one more question regarding the training of new model. In the fast-neural-style folder, their is a file called DataLoader.lua which contains following lines,

self.image_paths = {
train = '/train2014/images',
val = '/val2014/images',
}

Is that the train and val are paths to the training and validation images? This file also takes in input .h5 file which is already being created with the train and validation images.

I got following error,
/fast_neural_style/DataLoader.lua:38: attempt to call method 'dataspaceSize' (a nil value)

I think this is due to the path specified in self.image_paths. Please let me know if it is the solution.

[opedge]

Hi @theheena ,
I think DataLoader.lua is intended for training model. torch2coreml can convert only already trained model into CoreML for inference on mobile device.
I don't know exactly about internal structure of DataLoader.lua.