raise NotImplementedError("By default it is assumed that the data loader used for initialize "

[SiChuanJay]

Hello, I ran into a problem with the training-time quantization of yolov8 using nncf, as shown below. I refer to the example link (https://github.com/openvinotoolkit/openvino_notebooks/tree/main/notebooks/302-pytorch-quantization-aware-training). My v8_dataloader is also of type DataLoader, but I got an error. Thanks very much!

[eaidova]

@MaximProshin could you please help?

[MaximProshin]

@SiChuanJay , thanks for your interest! In general, we expect you to raise issues about nncf to nncf repo (https://github.com/openvinotoolkit/nncf), not here.

As far as I can see you're applying training-aware quantization for yolov8 while it can be successfully quantized in post-training. Can you please take a look at https://github.com/openvinotoolkit/nncf/tree/develop/examples/post_training_quantization/openvino/yolov8 or https://github.com/openvinotoolkit/openvino_notebooks/tree/main/notebooks/230-yolov8-optimization?

Regarding your error, I think it's due to the way you've implemented the dataloader. I let @vshampor provide more thoughts about it and how it could be fixed on your side.

[SiChuanJay]

Yes, I have tried the post-training method and it can be successful. But the train-time approach will fine-tune the model after it is converted to int8, which might work better, so I wanted to give it a try. Besides, I have another question. The int8 model I got using the example you provided is the same size as the fp32 model. Why?

[MaximProshin]

The issue has been moved to nncf repo: openvinotoolkit/nncf#1605. Please use that issue for further discussions.