Skip to content

A concise code for training and evaluating Unet using tensorflow+keras

License

Notifications You must be signed in to change notification settings

zizhaozhang/unet-tensorflow-keras

Repository files navigation

Unet-tensorflow-keras

A concise code for training and evaluating Unet using tensorflow+keras

A simple practice of the mixture usage of tensorflow and keras for the segmentation task. Sometime using Keras to manage the training is not flexiable. But we still want to utilize the convenience of Keras to build the model. Using Keras to build the model is super easy and fully compatible with Tensorflow.

I use the Unet architecture and modify its unsampling part to automatically adjust the feature map width and height when merge (concat) with previous layers. In this way, we do not need to compute the specific input size to fit the model but take an arbitrary size.

UPDATE, Oct 2017:

  • Take the image loader out of keras package to prevent the unpaird image/mask bug
  • Use Tensorflow.contrib.keras
  • Improve usability
  • Python 3 supported

UPDATE, July 2017:

  • Change the code to fit Tensorflow > 1.0
  • Adding a loader file to use Keras generator to load image and mask with automatic augmentation
  • Adding a VIS module to manage the evaluation metric.
  • Adding opt.py to support easier use

Usage

  • See loader.py to organize your train/test data hierarchy

  • Set necessary hpyerparameters and run train.py

    python train.py --data_path ./datasets/your_dataset_folder/ --checkpoint_path ./checkpoints/unet_example/
  • Visualize the train loss, dice score, learning rate, output mask, and first layer convolutional kernels per iteration in tensorboard

    tensorboard --logdir=train_log/
    
  • When checkpoints are saved, you can use eval.py to test an input image with an arbitrary size.

  • Evaluate your model

    python eval.py --data_path ./datasets/your_dataset_folder/ --load_from_checkpoint ./checkpoints/unet_example/model-0 --batch_size 1

About

A concise code for training and evaluating Unet using tensorflow+keras

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages