QAT NN tutorial

Zama Concrete ML
1

Hi,
I have been working through the 4 notebook tutorial cifar/cifar-brevitas-finetuning. In the 2nd notebook (CifarQuantizationAwareTraining) I got the following error when trying to plot the results (plot_baseline(param_c100, test_loader_c100, device):
TypeError: to() received an invalid combination of arguments - got (dict), but expected one of:

  • (torch.device device, torch.dtype dtype, bool non_blocking, bool copy, *, torch.memory_format memory_format)
  • (torch.dtype dtype, bool non_blocking, bool copy, *, torch.memory_format memory_format)
  • (Tensor tensor, bool non_blocking, bool copy, *, torch.memory_format memory_format)
    I have followed all the steps and generated the quant_vgg results without a problem. Have you got any idea why I might be getting this error? I am using a venv on a Mac and had to revert to torch 1.13 to use concrete-ml.
    Many thanks for the great step by step tutorial and all the tutorial videos.
3

It seems to be a bug:

In cifar_utils.py at line 258 the torch_inference function should only get 3 arguments.

It’s:

    baseline = torch_inference(fp32_vgg, data, param, device)

but should be:

    baseline = torch_inference(fp32_vgg, data, device)
4

Thank you Andrei! I was also wondering if Zama papers are presented at any particular conference? I would like to be up-to-date with all the latest developments.
Kind regards,
Helen

5

Check out our events page: Events | Zama! FHE.org is a big event that we usually submit papers to.

1 Like
6

And all our papers are on GitHub - zama-ai/awesome-zama: A curated list of amazing Fully Homomorphic Encryption (FHE) resources created by the team at Zama.! Cheers

7

Thank you, Benoit. This is awesome! I am enjoying investing Zama’s work and following the research.
Many thanks,
Helen

1 Like