Unofficial code of MSRF-Net developed in PyTorch
------------ IN PROGRESS ------------
- Write the model code based on official TF code.
- Write the training/evaluation code.
- Improve the training/evaluation code adding some stuff to tensorboard.
- Write the test code.
- Write the inferencing (demo) code.
- Train and test.
- PyTorch 1.9.0 was used with cuda 11.1.
- The hyperparameter init_feat was added. It controls the number of initial channels for the UNet. In the original code It was 32. I recommend to use a power of two because the reduction ratio in Squeeze and Excitation blocks.
- The Shape Stream isn't copy exactly from official code, It was copied from the original Shape Stream repo.
- Added image visualization during training to TensorBoard. This improvement will help you to check the performance during training.
During training DICE coefficient (in loss and as a metric) is computed without the BG.
