Add fp16 PyTorch models

[kkontny]

In Pytorch you can convert model to fp16 with module.half() call. I think it should be called before converting to TorchScript. See docs https://pytorch.org/docs/stable/generated/torch.nn.Module.html. I think it should be quite simple implementation.

[jan-grzybek-ampere]

@dkupnicki please take a look

[dkupnicki]

I added self.__model.half() before

ampere_model_library/utils/pytorch.py

Line 46 in 209549f

self.__frozen_script = torch.jit.freeze(torch.jit.script(self.__model))

and tested a few models. All of them stopped working, but they threw slightly different errors:

• resnet_50_v1 and a few others threw RuntimeError: "rsqrt_cpu" not implemented for 'Half'
• alexnet and ssd_vgg_16 threw RuntimeError: expected scalar type Float but found Half
• roberta_base_squad threw RuntimeError: "LayerNormKernelImpl" not implemented for 'Half'

Looks like half() only works on GPUs.

[kkontny]

Yes, it is expected. Since x86 doesn't support fp16 natively, nobody cared about this on CPU. However Altra supports it natively, so when eager mode will be implemented I expect it to work.