This is small project, for study U-Net-style architecture
For this project I used Oxford pets dataset. This dataset has 7930 images of dogs and cats, with mask for segmentation training.
Model is simply U-Net-style network, with encoder based on convolution-relu-batchnorm block, and max pooling, and decoder with deconvolution-convolution and concatenation step.
I trained this network on Kaggle kernel with Tesla P100 GPU. After 15 epochs, results are pretty neat.
