This is a simple implementation of the U-Net arhitecture and a project to utilize segmentation using Tensorboard.
A Korean translation of U-Net Paper: http://bit.ly/UNet_Paper_Translation
This tutorial depends on the following libraries:
- Tensorflow == 1.14
- Keras == 2.3.1
My computing resources are as follows:
- CPU: Intel i7-8700k
- GPU: GTX 1080ti
- RAM: 64GB
python main.py
Tensorboard
tensorboard --logdir=./logs --host localhost
The original dataset is from ISBI challenge, and I've downloaded it and done the pre-processing.
You can find it in folder data/membrane.
- Data Augmentation
10 times more images were used from the original number.
Method | Value |
---|---|
Rotation Range | 0.2 |
Width Shift Range | 0.05 |
Height Shift Range | 0.05 |
Shear range | 0.05 |
Zoom range | 0.05 |
Horizontal Flip | True |
Fill Mode | reflect |
- Hyperparameters
- epochs : 50
- batch size : 5
- Learning rate : 0.0001
- Optimizer, Loss function and Metric
- Adam
- Binary Cross Entropy
- Accuracy
Model performance was approximately 91% accuracy for validation data when 50 epochs were trained.