Pipeline for U-Net FCN with Keras and Tensorflow backend
U-Net is an FCN that is popular for biomedical image segmentation. This repository includes the pipeline for preprocessing, training, and testing images.
This was successfully run in a Conda environment with Python 3.7
Commands below are meant to be used in Linux terminal.
Create environment from Conda Base environment:
cd U-Net-FCN
conda create -n unetenv python==3.7
Install pip in conda environment:
conda install pip
Install requirements and unzip directories:
python imports.py
To train model run unet_train.py and follow prompts:
python unet_train.py
Open Tensorboard to monitor loss and Jaccard coefficient across epochs:
tensorboard --logdir=logs/scalars/ --reload_multifile True --reload_interval 5
To test model run test_unet.py and follow prompts:
python test_unet.py
To convert h5/hdf5 to yaml for much smaller file size, yet slightly lower accuracy, run h5_to_yaml.py and follow prompts:
python h5_to_yaml.py
Weights file (.hdf5/.h5) can also be reduced by pruning model model during training with tensorflow_model_optimization.sparsity (see Tensorflow Model Optimization Pruning).
U-Net FCN Architecture
