Printing Quantized Model Weights

quantization
1

How do we print quantized model weights in PyTorch?

To print using normal PyTorch representation, I understand we use the following approach…

print_parameters = lambda model: [print(name, param.data) for name, param in model.named_parameters() if param.requires_grad]

Similarly, if I defined a model as follows…

import torch

class Model(torch.nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.conv1 = torch.nn.Conv2d(2, 3, 1, bias=False)
        self.conv2 = torch.nn.Conv2d(3, 1, 2, bias=False)

    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        return x

model = Model()

Then we can print using pretty straight-forward syntax…

print(model.conv1.weight)
print(model.conv2.weight)

However, both of these approaches fail when the model is converted to a quantized form. Specifically, after the following procedures…

model.qconfig = torch.quantization.default_qconfig
torch.quantization.prepare(model, inplace=True)
torch.quantization.convert(model, inplace=True)

Printing model.conv1.weight returns a method and the loop provided at the beginning does not print anything.

2

We have recently prototyped PyTorch Numeric Suite that can compare the numerics between float and quantized model. It’s in nightly build and will be in 1.6.

You can take a look at the following example to see how to compare the weights using it: