Skip to content

karl-schulz/deepdecoder

 
 

Repository files navigation

RenderGAN

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

You might be interested in:

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

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 94.8%
  • Makefile 2.7%
  • Dockerfile 1.7%
  • Shell 0.8%