boliu-svcl / meta-open Goto Github PK

what i get from the paper is
(1) there is a stage-like train procedure
(2) there is a h like meta-model to aux the backbone.
(3) there seems a unseen classes in the test.

what i get from the code is
(1) the loss-function could consists three parts.
i. the few-shot loss-functions, aka, the -Euclidean distance.
ii. the open-set loss-functions, aka, the papers' open-set evaluation algorithm.
iii. the aux part. this part use the fc to evaluation logits. is the fc the paper's h-thing?
(2) the dataloader gather the same-class for few-shot ,open-set and aux types, but different images from the train-set. could it be possible to the unseen classes?

thanks a lot.

In original setting, the backbone is resnet10. I try to change it to resnet18, so in 'default.yaml', I change the parameter model.structure: 'resnet10' to model.structure: 'resnet18', while other parameters unchanged. But the result I got is much lower than the result in the paper. Is there something wrong with my operation？

Hi, thank you for sharing your experimental code.

In the paper, the OpenMax baseline uses the negative of distances like ProtoNet.

But, the lower $\omega$ value in OpenMax indicates the query sample is more likely to be an unknown sample.
Therefore, I think the OpenMax method assumes all elements in the activation vector to be non-negative.
However the logits of ProtoNet are non-positive, which is inconsistent with the assumption.

Also, since there is a single sample for each class in 1-shot cases, it is nearly impossible to fit the Weibull model using supports as stated in the OpenMax paper (e.g. The OpenMax code uses 20 samples for each class to fit the Weibull model).

Due to those problems, I could not reproduce your result on the baselines.

Could you give more detailed information to reproduce your baselines (Openmax & Counterfactual)?

Thank you!