How to do transfer learning with inception_v3?

1

Since very recently, inception_v3 is available in torchvision.models and getting it is as simple as typing model = models.inception_v3(pretrained=True)

Since the model was pretrained on ImageNet it has 1000 output classes which I wanted to change to 2 for my binary classifier - so I created the following head:

from collections import OrderedDict
head = nn.Sequential(OrderedDict([
                          ('fc1', nn.Linear(2048, 1024)),
                          ('relu1', nn.ReLU()),
                          ('fc2', nn.Linear(1024, 512)),
                          ('relu2', nn.ReLU()),
                          ('fc3', nn.Linear(512, 128)),
                          ('relu3', nn.ReLU()),
                          ('fc4', nn.Linear(128, 32)),
                          ('relu4', nn.ReLU()),
                          ('fc5', nn.Linear(32, 2)),
                          ('output', nn.LogSoftmax(dim=1))
                          ]))

My question is regarding how to add this head. If I was using restnet54, this would be as simple as model.fc = head but in the case of inception_v3 I have realized there is what it looks like as two output layers:

image
image809×246 18.6 KB

and by doing model(inputs) I also get and InceptionOutputs(logits=tensor([[-0.8333, -0.5702] [...], device='cuda:0', grad_fn=<LogSoftmaxBackward>), aux_logits=tensor([[ 1.0039, 2.4251] [...]

Do I need to change both heads? What is the difference between logits and aux_logits. I see the term ‘logits’ being used very often but I struggle to make sense of it (i.e. I understand the difference between scores and probs after a softmax, not sure where logits and aux_logits fit in this schema).

This would affect the whole training process, i.e. it can no longer be only optimizer = optim.Adam(model.fc.parameters(), lr=0.001)

2

inception_v3 was added in March 2017, so I’m not sure, if we are talking about the same model.

Anyway, the Finetuning Tutorial has specific code paths for the inception model and for its auxiliary outputs.

Logits are basically the input to a softmax layer.
Since nn.CrossentropyLoss is using F.log_softmax and nn.NLLLoss internally, you don’t need to apply softmax manually.

3

Thanks @ptrblck! I was coding super late and I got confused with EfficientNet :face_with_hand_over_mouth: but your link and explanation was super useful! Thanks again.