At finetue stage, is that classifier is new, not finetune from trained?

[banjuanshua]

Thank's for your work! I have a question that is it classifier at test stage is new? I can't find classifier loading trained weights in code.

[wyharveychen]

The classification layer of baseline and baseline++ is new. For other meta-learning methods, protonet/matchingnet/relationnet do not have classification layer, and MAML would load the trained weight for the entire model (including classification layer).

[banjuanshua]

Tanks for your answer. And did you analyze the difference of protonet implement be between yours (74.65% ResNet34) and MetaOptNet(75.60% ResNet12 feature_dimension=16000). I find backbone is different. What's the feature dimension in your protonet?

[wyharveychen]

Hello, for the ResNet 34 model, the feature dimension of the final layer is 7x7x2048, but we have included the avg pooling layer, so our feature dimension is 2048.

[wyharveychen]

As the difference is not large, it could depend on parameters for training.