Pytorch implementation for paper Renet (Relational embedding network).
- Python==3.9.6
- torch==1.7.1
- torchvision==0.8.2
- GPU: NVIDIA GeForce RTX 3090 * 1
-
Self-correlation Representation: a base feature map -> a self-correlation tensor
-
Cross-correlation Attention: a self-correlation tensor -> co-attention
According to the split in ICLR(Meta-Learning with Fewer Tasks through Task Interpolation), I split the raw mini-Imagenet dataset to reduce its training classes by specific sequence.
According to the split in ICLR(Meta-Learning with Fewer Tasks through Task Interpolation), I split the raw Dermnet dataset to reduce its training classes by specific sequence.
ISIC 2018 is an extremely unbalanced dataset which only has 7 classes in total. So I trained it in 2-way 5-shot mode, while others are trained in 5-way mode.
According to the training process in ICLR(Meta-Learning with Fewer Tasks through Task Interpolation), I remove the validation process during every epoch of training.
| Renet | Renet(gamma=2) | ProtoNet | MAML | |
|---|---|---|---|---|
| ISIC 2018 | 78.033 | 64.245 | 78.180 | 61.42 |
| Mini-Imagenet-S(12) | Mini-Imagenet-S(25) | Mini-Imagenet-S(38) | Mini-Imagenet-S(51) | Mini-Imagenet(64) | |
|---|---|---|---|---|---|
| Renet | 65.231 | 73.133 | 76.521 | 79.226 | 80.463 |