So u just provide the code for stage B?

[abcdvzz]

No description provided.

[phecy]

for the pre-train stage, we just use standard AMDIM for self supervised training.https://github.com/Philip-Bachman/amdim-public

And the we released some pretrained models.

[semi-supervised-paper]

What's the hyper-parameters of AMDIM for Mini-ImageNet? And what's the data augmentation of Mini-ImageNet in AMDIM?

[phecy]

@semi-supervised-paper we use AmdimNet(ndf=192, ndepth=8, nrkhs=1536) to train Mini-ImageNet. For more detail:

python train.py \
  --ndf 192 \
  --n_rkhs 1536 \
  --batch_size 480 \
  --tclip 20.0 \
  --n_depth 8 \

[semi-supervised-paper]

Thanks for your kindly response. I wonder if you have tried other powerful unsupervised learning methods?

[phecy]

we do not do experiment with other ssl algorithms, we believe that other ssl algorithms which aim to learn task-agnostic representations can improve the performance of few-shot learning task.

[abcdvzz]

Could u pls give us a detailed configuration of the training of the CUB dataset on stage A? For example, input size, epoch, and transformation. I only get an accuracy of 17% following the Imagenet128 configuration. Thank u.

[jaemyungkwon]

Hi, i want to know this code is ssl with meta learning?
and. do you know how to use multi-gpu for your code?