No compression_loss needed for QAT?

[EmonLu]

Hi,

I noticed there're two Usage.md under docs/usage/training_time_compression, one for QAT and one for other algorithms.

1. Why there is no compression_loss in QAT docs? How does quantization update the parameters like scalefactor and zeropoint?
2. Is PTQ indispensable in QAT?
3. Are there any tips for QAT to have a better performance? e.g. use PTInitialDataloader to initialize

[alexsu52]

Hello @EmonLu,

At the first, I would lke to note that create_compressed_model(...) API will be deprecated in the next release and nncf.quantize is recommended API for QAT.

Why there is no compression_loss in QAT docs? How does quantization update the parameters like scalefactor and zeropoint?

QAT uses the model loss that is used for training and does not require additional compression losses.

Is PTQ indispensable in QAT?

PTQ provides good initialization for QAT to have a better performance.

Are there any tips for QAT to have a better performance? e.g. use PTInitialDataloader to initialize

PTInitialDataloader will be deprecated in the next release as well. Please take a look at Jupyter notebook on Quantization Aware Training with NNCF and PyTorch which demonstrates how to use NNCF API for QAT and includes general tips to achieve a good performance.