Alexander Kozlov

I would even remove a connection to any numpy version from the user's requirements. Internal validation can use a fixed version for reproducibility. @vshampor, @alexsu52.

Option 3 seems preferable but it should cost some substantial effort I guess. Anyway, I vote for it.

Hi @xiaoyaopeng, OpenVINO does it for you automatically for floating-point and quantized models. And there is no accuracy degradation if you use NNCF.

This can be a problem actually because OpenVINO mostly fuses BN into FakeQuantize parameters. To be honest of the HW we have supports per-channel quantization of weights as the most...

NNCF can export quantization parameters either to standard ONNX with QuantizeLinerar and DequantizeLinear operations or to custom ONNX with FakeQuantize op from openvino domain. Both are recognizable by OpenVINO. As...

@kblaszczak-intel, can we assign anybody to review this and integrate it into the documentation of the upcoming release? This is actually a back-merge of the reviewed [PR](https://github.com/openvinotoolkit/openvino/pull/12203) from master.

@kblaszczak-intel, we need to have these docs in OV 2022.2. Can anybody integrate them? cc'ed @tsavina.

Hi @korotaS, Can you please share the NNCF config?

Thanks, @korotaS! This looks like an absence of the specific fusion pattern in the NNCF PT version. Is it possible for you to share the quantized and not quantized models...

> @AlexKoff88 of course: > > ```python > nncf_config_dict = { > "input_info": {"sample_size": [1, 3, 512, 512]}, > "log_dir": 'notebooks/nncf_logs/', > "compression": { > "algorithm": "quantization", # specify the...