CrossTransformers-PyTorch

Unofficial PyTorch implementation

• CrossTransformers: spatially-aware few-shot transfer
  • Carl Doersch, Ankush Gupta, Andrew Zisserman
  • NeurIPS 2020

Main contributions

• CrossTransformers Architecture
• SimCLR episodes

Implementation details

• Resnet34, output feature map 14x14 by using dilated conv
• Higher image resolution (224x224)
• Strong data augmentation following [2]
• Normalized gradient descent
• 50% episodes of uniform category sampling
• First step: Pretraining feature extractor on train categories, early stop by linear classifier accuracy on validation categories.

TODO

• CTX sanity check on miniImagenet
• CTX on Meta-Dataset [1]
• CTX + SimCLR Eps
• CTX + SimCLR Eps + Aug

Acknowledgements

• miniImagenet experiments based out of DN4 codebase.

References

[1] Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples.

[2] Optimized generic feature learning for few-shot classification across domains.