Repository for few-shot image translation applications, in particular SEVIR but also miniimagenet for fast prototyping. Dataloaders and patch discriminator are borrowed from here, the UNet model architecture from here, and MAML gradient calculations from here.
- test pretrained backbones from multiple epochs
- test new pretraining
- do not adapt discriminator
- plot discriminator losses throughout training
- improve bookkeping: create results dir; use os.path.join()
- train joint reconstruction
- train MAML reconstruction
- train joint GAN
- train MAML GAN
- eval joint reconstruction
- eval MAML reconstruction
- eval joint GAN
- eval MAML GAN
- load pretrained encoder
- save opt to disk
- sweep number_inner_steps for sevir
- test on MAML (e.g. implement inner loop in eval_unet)
- test contrastive pretraining after limiting train set
- add miniimagenet dataloader
- MAML + pretraining
- redo SEVIR dataset