DenoiSeg
DenoiSeg copied to clipboard
Joint training of denoising and segmentation.
From @tibuch in https://github.com/juglab/DenoiSeg/pull/40: The first epoch usually takes super long. In general it feels not very performant.
From @tibuch in https://github.com/juglab/DenoiSeg/pull/40: The image segmentation tensorboard is currently not working and commented. The normal loss tensorboard is available.
From @tibuch in https://github.com/juglab/DenoiSeg/pull/40: Datasets are currently limited, which leads to problems if steps_per_epoch x batch_size is larger than the dataset length. As long as steps_per_epoch x batch_size is smaller...
From @tibuch, modifying the loss so that the masking is per pixel instead of per patch. I drop the idea here otherwise it might get forgotten.
DenoiSeg requires `keras` 2.2.5 (as per README instructions, but also during `pip install denoiseg`). This version of `keras` contains a workaround for a specific `h5py` version, where dataset attributes are...