/SwinIA-segmentation

Exploring emerging segmentation abilities of SwinIA models

Primary LanguagePython

Noise2Same (PyTorch)

PyTorch reimplementation of Noise2Same. Work in progress.

Usage

Default configuration is located in config/config.yaml. Experiment configs config/experiments may override defaults.

Training

To run an experiment for BSD68, execute

python train.py +experiment=bsd68

Four experiments from Noise2Same are supported: bsd68, hanzi, imagenet, planaria.

Training logs and model weights will be saved to resuts/train/datetime.

Evaluation

To run evaluation for BSD68, execute

python evaluate.py +experiment=bsd68

By 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.

Results replication

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