explicit_class_member_call.md

Avoid explicit Invocation of Custom Class Members Outside the Class

This restriction also comes from the implementation of torch.fx(to do the quantization). torch.fx will translate our torch.nn.Module into torch.fx.GraphModule, which is a torch.nn.Module instance that holds a Graph as well as a forward method generated from the Graph. During this process, all the self-defined class members will be discarded. Two cases is listed below:

• The first is also from the cellpose repo. The code snippet is here.

if self.pretrained_model:
   self.net.load_model(self.pretrained_model[0], device=self.device)

In this case, the self.net is a torch.nn.Module instance, and is the target for our quantization function. During the calibration process in the quantization, the self.net will be transformed to the torch.fx.GraphModule and will cause an error when the self.net.load_model is called.

• The second cases come from the Deconoising repo. The code snippet is from here

def predict(im, net, device, outScaling):
    stdTorch=torch.Tensor(np.array(net.std)).to(device)
    meanTorch=torch.Tensor(np.array(net.mean)).to(device)
    ...

In this example, we use this function as our calibration function. However, the net is a torch.nn.Module instance, which will be transformed to torch.fx.GraphModule and cause an error when the net.std is called. We can precomputed these values.