Quantized Pre Trained Model Urls

quantization
1

Hi,

I found this repo https://github.com/pytorch/vision/blob/release/0.8.0/torchvision/models/quantization/googlenet.py containing quantized versions of models like GoogleNet and ResNet. I see that model weights are loaded from the URLs.

quant_model_urls = {
    # fp32 GoogLeNet ported from TensorFlow, with weights quantized in PyTorch
    'googlenet_fbgemm': 'https://download.pytorch.org/models/quantized/googlenet_fbgemm-c00238cf.pth',
}

I have the below question on these weights:

  1. What quantization type is used for these weights, for example, QAT or Post static quantization?
  2. Do we have such URLs for other torchvision models like Densenet and VGG16?

Thanks

2

@Sairam954 I believe there is a method here: vision/utils.py at release/0.8.0 · pytorch/vision · GitHub
It might give you some idea on how the models are quantized.

@jerryzh168 @raghuramank100 Is there more info on these? SHould we put it in the docs?

3

  1. here is some additional information: vision/README.md at release/0.8.0 · pytorch/vision · GitHub, it should be post training static quantization I think.

  2. looks like we didn’t quantize these networks right now, but feel free to try quantizing them with FX graph mode quantization api ((prototype) FX Graph Mode Post Training Static Quantization — PyTorch Tutorials 1.11.0+cu102 documentation), or you can wait a few months, we will uploading the weights for some reference models that are quantized with fx graph mode quantization as well. cc @andrewor