API for torch.load()

[iprovalo]

I need to be able to read a serialized shared non-layer specific tensor (weights) which is saved like so: torch.save(model.shared.weight, 'weights.bin')

In Python it can be read: torch.load('weights.bin')

Is there a way to load this tensor in DJL?

Thank you!

[frankfliu]

We only support load numpy format, you can convert it to numpy first and save it in either .npy or .npz format.

[iprovalo]

Thank you, @frankfliu ! What is the API for that?

[frankfliu]

You can use:

NDArray.decode(NDManager, byte[])
NDManager.decode(byte[])
NDManager.decode(InputStream)

for .npz:
NDList.decode(NDManager, byte[])
NDList.decode(NDManager, InputStream)

See examples: https://github.com/deepjavalibrary/djl/blob/master/integration/src/main/java/ai/djl/integration/tests/ndarray/NDArrayOtherOpTest.java#L821

https://github.com/deepjavalibrary/djl/blob/master/api/src/test/java/ai/djl/ndarray/NDListTest.java#L26

[iprovalo]

Thank you again, @frankfliu !