karl-schulz/deepdecoder

RenderGAN

This repository holds the code to the "RenderGAN: Generating Realistic Labeled Data" paper.

You might be interested in:

• implementation of the RenderGAN
• code of the generator network
• discriminator
• architecture of the decoder network
• the Makefile to orchestrate the training process
• the hyperparameters for the training

Run the code

1. Build the docker image as described in docker/README.md

2. Start a docker container and ensure you have enough disk space ~300GB.

3. Mount or clone this repository into the docker container. Let $REPO_DIR contain the path to the repository.

4. Install the repository with pip3 install -e . .

5. Download all files from this google drive directory and place them in the $REPO_DIR/data directory. You might find the gdrive cli tool useful. Check the sha1 sums:

   db00d1ae8d1920c6d4faa830cebcd5783b44e905  gt_test_shuffled.hdf5
   37961551dd062179cafc3f7eee3a75e113555c59  gt_train_shuffled.hdf5
   be1870252dd926aeff8818ad13d9b28a1253935e  real_tags.hdf5
   

6. Now your can run the code.

   $ mkdir build
   $ echo "{\"path\": \"`realpath data/real_tags.hdf5`\"}" > build/real_dataset
   $ realpath data > build/gt_data_dir
   $ cd build
   $ bb_make ../config/train tag3d_network

   The bb_make program is a s lim wrapper around this Makefile. The first argument is the directory of the parameters you want to use and the second is the make target to execute.

   Targets:

   • tag3d_data_set: Generates a dataset using the simple 3D model.
   • tag3d_network: Train a neural network to emulate the simple 3D model.
   • rendergan: Trains the GAN.
   • artificial_data_set: Samples an artificial dataset from the RenderGAN.
   • decoder_model: Trains a neural network to decode the tags.
   • decoder_evaluation: Evaluates a trained decoder network and prints the mean hamming distance.
   • decoder_on_real: Train a neural network on the limited set of real data.

   Every target has a corresponding setting file (e.g. setting the number of sample to generate ...). The decoder_model target can be run with different settings using the -o flag:

   $ bb_make ../config/train -o decoder_settings=decoder_handmade_tag3d.yaml decoder
   

   This will execute the decoder target but with the decoder_handmade_tag3d.yaml file.

   You can also copy the settings directory $ cp -r config/train config/my_settings and experiement with your own hyperparameters. Then you have to use $ bb_make ../config/my_settings