PyTorch reimplementation of Noise2Same. Work in progress.
Default configuration is located in config/config.yaml.
Experiment configs config/experiments may override defaults.
To run an experiment for BSD68, execute
python train.py +experiment=bsd68Four experiments from Noise2Same are supported: bsd68, hanzi, imagenet, planaria.
Training logs and model weights will be saved to resuts/train/datetime.
To run evaluation for BSD68, execute
python evaluate.py +experiment=bsd68By default, we assume the weights for the model to be in weights/experiment.pth
but you can specify the path by adding +checkpoint=/path/to/checkpoint.
Model's outputs and scores (RMSE, PSNR, SSIM for each image) will be saved to resuts/evaluate/datetime.
We replicate the main results of Noise2Same (Table 3)
| Dataset | Ours (Noise2Self) | Noise2Same paper | Ours (Noise2Same) | Weights |
|---|---|---|---|---|
| BSD68 | 26.73 | 27.95 | 28.11 | Drive |
| HanZi | 14.38 | 14.83 | Drive | |
| ImageNet | 22.26 | 22.81 | Drive | |
| Planaria (C1/C2/C3) | 29.48 / 26.93 / 22.41 | 29.14 / 27.11 / 22.80 | Drive |