Initializing QAT with pre-trained quantization parameters

quantization
1

Hi all,

I apologize if this question is covered elsewhere.

I would like to perform quantization-aware training, but with the model initialized according to the pre-trained, post-training-quantized quantization parameters (e.g., a torchvision quantized model with layers initialized with the same scale, zero_point, etc. as in the pre-trained quantization model that is initialized with quantize=True).

That is, I’d like the initial model used for QAT to produce the same output as a pre-trained model that has been quantized using a post-training method (e.g., static quantization).

Is there an easy way to achieve this? I had tried hacking manually setting some of the QAT model’s FakeQuantizer parameters, but was unable to get it working properly.

I appreciate any help! Please let me know if my question is unclear and I will rephrase it.

Thanks!

2

I don’t think that is supported right now, instead if you have access to the original floating point model you can just do qat with that. you’ll get the model with same accuracy as the post training model if you call prepare_qat for the original model and calibrate it with the same data.

3

Got it. Thanks for the response!