Graph Mode Quantization Aware Training in PyTorch

quantization
1

Hi,

The Graph Mode Post Training Quantization in Pytorch is awesome because it doesn’t need the model to be defined in certain way only using Modules. (https://pytorch.org/tutorials/prototype/graph_mode_static_quantization_tutorial.html)

Wondering if there is a plan for Graph Mode Quantization Aware Training as well in Pytorch?

Thanks,
Manu

2

Hi Manu,

Yes. It’s very much work in progress and the code is at https://github.com/pytorch/pytorch/tree/master/torch/quantization/fx Tutorials and docs will be released once it’s ready.

3

You can monitor the progress here: https://github.com/pytorch/pytorch/issues/45635

4

Thanks Daya,

What would be the rough timeline for graph mode QAT to be available - will it be included in PyTorch 1.7 release?

Also wondering what would be the difference between quantize_jit and quantize_fx

5

It’s not a part of 1.7. quantize_jit is the current api that you used for graph-mode post-training quantization and it will be deprecated once quantize_fx is available (i.e., new graph mode).

1 Like