nn.Linear, nn.Conv2d, nn.ReLU,
nn.LeakyReLU, nn.Sigmoid, nn.Tanh,
nn.Softmax, nn.Flatten, nn.BatchNorm1d/2d,
nn.Dropout, nn.MaxPool2d, nn.AvgPool2d,
nn.CrossEntropyLoss, nn.MSELoss, nn.L1Loss,
nn.SmoothL1Loss, nn.BCELoss, nn.BCEWithLogitsLoss,
nn.NLLLoss, nn.KLDivLoss. Unsupported layers are skipped.
Paste a whole class Net(nn.Module) and the layers are ordered by
how forward() calls them (functional ops like F.relu,
F.max_pool2d, x.view(…) included).