Is there any problem with using "sub models"?

1

I wish to create a model which uses a resnet as part of the architecture

It will eventually have other layers, but I’ve put together a quick prototype:

class CombinedModel(torch.nn.Module):
    def __init__(self):
        super(CombinedModel, self).__init__()

        self.resnet = torch.hub.load('pytorch/vision:v0.10.0', 'resnet101', pretrained=True)

    def forward(self, x):
        x = self.resnet(x)

        return x

This seems to learn fine with my (very standard) training code. Are there any things I need to be aware of though? Do I need to forward, for example, .eval and .train calls to the sub-model?

2

Will depend if the resnet is already trained with weights that are “good” for your data For the evaluation part you will definitely have to call to sucessfully run it.

3

No, all properly registered submodules will automatically receive the .train(), .eval(), and .to() calls and you can double check it:

model = CombinedModel()
print(model.resnet.training)
# True

model.eval()
print(model.training)
# False
print(model.resnet.training)
# False
2 Likes