A small project to encode Spanish politicians' pictures into 2 dimensions and build a latent map.
The objectives of this project are:
Show that all congress' members look alike, despite their affiliation.
Explore the use of different virtualization solutions in Python, that could take advantage of a native GPU despite the virtualization layer. A priori, one option is pipenv, to control both dependencies and environment. Another option is docker.
Also use Keras as interface for creating neural networks. The idea is training a autoencoder model, storing it and applying it later to create a visual result.
Use pipenv to install and open a shell. You can use it too to check code style with pipenv run style. Unluckily this will not use your GPU.
If you have a machine with a NVIDIA GPU, using a docker container is the only virtualization option that will use the GPU.
You need to have both docker and NVIDIA drivers installed first. Then install nvidia for docker in your machine, and you will be able to use any tensorflow docker container. Look for the correct tag, like version-gpu, for instance 2.1.0-gpu.
There is a Dockerfile ready to work with. Just create the container with:
docker build -t all-are-the-same .
Later use it to execute anything, like:
docker run -t -v $(pwd):/app all-are-the-same python 2.train_autoencoder.py
- Keras : Creating an autoencoder is extremely simple with Keras.
- Keras : Using MSE makes decoded images too blurry, so it'd better to use other error scoring (like perceptuals).
- pipenv : Some of the Cuda libraries need by tensorflow to use a GPU are not possible to install in a virtual environment (virtualenv or pipenv), so in order to use a GPU in a virtualized environment, docker is the only easy solution.
- docker : The only option to virtualize a project that uses a local GPU.
- Update to tensorflow 2
- Explain some lessons learned
- Use docker
- Complete the code for the 2-dim map
- Try other loss functions, like perceptual ones, to avoid blur