Place minmaxobserver on activation layer

quantization
1

When I apply default qconfig to my model, it automatically puts minmaxobserver on conv and batchnorm layers, but not activation layers. But I need to get the min/max values of those layers too. Is there a simple way to do that?

import torch.ao.quantization as Q
testmodel = nn.Sequential(
    nn.Conv2d(10,10,3),
    nn.BatchNorm2d(10),
    nn.PReLU(10),
)
testmodel.qconfig = Q.default_qconfig
Q.prepare(testmodel, inplace=True)
print(testmodel) # no observer on prelu
2

What’s the activation layers refer to? nn.PRelu?

3

Yes. If it was relu or sigmoid I could infer the value range from the batchnorm output, but that’s not the case for prelu

4

I found that some activation modules (including nn.Hardtanh, nn.Hardswish) get an observer, but others don’t. I’m not sure where the difference lies.

5

It seems to has to do with the function
torch.ao.quantization.quantization_mappings.get_default_qconfig_propagation_list(),
which lists all quantizable modules. I can override it by specifying the allow_list in torch.ao.quantization.prepare()