pix2pix from scratch
An attempt at creating a pix2pix implementation from scratch, using only Tensorflow as a dependency.
This largely follows Tensorflow's pix2pix tutorial:
https://www.tensorflow.org/tutorials/generative/pix2pix
Image Preparation
Images are required to be 1024x512, where the left image is the "real" image and the right image is the input image. You can use Figment to prepare these :-)
Setup
I use a machine with Ubuntu 20.04 and Lambda Stack.
The easiest method was to use Docker:
docker run -it --rm -v $PWD:/tf -w /tf -p 8888:8888 --gpus all tensorflow/tensorflow:latest-gpu-jupyterThen open a browser to the shown URL and open pix2pix.ipynb.
Change the directories as required to set up your project.
Convert Model
Once the model is running, run the last cell to convert the generator to Keras (.h5).
To convert the saved model to tensorflowjs:
pip install tensorflowjs
tensorflowjs_converter --input_format keras generator.h5 tfjs/model/