UNet++

A PyTorch implementation of UNet++: A Nested U-Net Architecture for Medical Image Segmentation trained with CelebAMask

Dataset: CelebAMask

I made new datasets in Kaggle after pre-processing.

1. Use only background for mask images
2. Downsize to 128x128 & 256x256

Implementation Details

optimizer: Adam
learning rate: 3e-4 (without scheduler)
image size: 128x128
epoch: 100
batch size: 12

train loss: 0.5 * binary cross entrophy + dice coefficient loss
validation & test metric: IoU (Intersection over Union)

Output of test images

Some outputs are even better than target images in the dataset.

Train Loss

IoU

	Train	Valid	Test
# of images	27000	2500	500
IoU	0.9892 (Epoch 100)	0.9616 (Epoch 93)	0.9618

Convert GIFs