Pytorch pix2pix implementation

How to run ?

• cd in this directory
• Start a docker env for pytorch sudo docker run --rm -it --init --runtime=nvidia --ipc=host --user="$(id -u):$(id -g)" --volume=$PWD:/app -v /your/db/folder:/data -e NVIDIA_VISIBLE_DEVICES=0 anibali/pytorch /bin/bash
• Install requirements via pip install -r requirements.txt
• You can run the training via python3 run_training.py

Dataset modes

• Colorisation
• Edge2Something (see examples, edge to Delaunay or edge to ADE20k)
• Labels2ADE20k

Using tensorboard to visualize the training

Tensorboard is a great tool and is wonderful in this case to see where your training is going. You can run sudo docker run -d -p 6006:6006 -v $(pwd)/logs:/logs --name my-tf-tensorboard volnet/tensorflow-tensorboard in the directory of training and will get nice visualisation.

For the ADE20k edges dataset, the training visualisation helps to see when the system stagnates.

In this example, delaunay paintings texture are learned from the edges

What can I get ?

Mostly bad modern art for the edge to Delaunay

The training is rather difficult for the edge to something, another example on ADE20k (after 130 epochs, trained on 512px images, batch size 2, two discriminators, lr=2e-4)