Bad Performance with default_observer for quantization

❓ Questions and Help

Please note that this issue tracker is not a help form and this issue will be closed.

We have a set of listed resources available on the website. Our primary means of support is our discussion forum:

• Discussion Forum

I tried to quantize mobilenet v2 from float model file.
I found defaut_observer is used for activation in QConfig.
https://github.com/pytorch/vision/blob/master/torchvision/models/quantization/utils.py#L27

I got bad imagenet classification accuracy with this configuration.

https://pytorch.org/docs/stable/quantization.html
Here, the following configuration is recommended.
qconfig = torch.quantization.get_default_qconfig('qnnpack')

this uses HistogramObserver.
qconfig = QConfig(activation=HistogramObserver.with_args(reduce_range=False),
weight=default_weight_observer)

When I changed default observer to HistogramObserver, I got much better accuracy.

I think the following configuration should be changed as recommended in official docs.
https://github.com/pytorch/vision/blob/master/torchvision/models/quantization/utils.py#L27

Start at torchvision/models/quantization/utils.py around line 27 and compare its activation observer with the documented qconfig using HistogramObserver. Check the MobileNetV2 quantization configuration and the linked quantization documentation, then verify that the observer change produces the expected accuracy improvement.