Is it possible to convert a GPU pre-trained model to CPU without cudnn?

[shamangary]

Since the limited computer resource, I only have a CPU-only computer. However, some pre-trained model is based on GPU.

It seems that the torch installation will not contain cudnn when you don't have GPU on your computer. Therefore, I have no cudnn class in my torch. Moreover, there is no cudnn to nn conversion in the README.md.

Is there a way to convert a GPU pre-trained model to a CPU pre-trained model without using cudnn.torch?

[notimesea]

cudnn.convert(model, nn) works fine at least when you have gpu.

[adroit91]

We have observed that converting using cudnn.convert doesn't work for all modules, for example cudnn.ClippedReLU doesn't get translated into nn despite mention of API compatibility.

[roywei]

cudnn.TemporalConvolution is also not converted, since it's a wrapper on cudnn.SpatialConvolution.

[Mina1368]

Hi could you find a way?

Thanks, Mina

[NightFury13]

@notimesea Does the cudnn.convert() method work if we convert a GPU-based model to a CPU-based model using a machine which has GPU support and then load this converted model on a CPU-only machine?

[abarya]

@soumith, please help. I'm having the same error. I don't know why when I call graph.dot(final_model.fg, 'model', 'model') torch crashes.