self-driving-AttGAN

• Unofficial Tensorflow implementation of Attribute-Controlled Traffic Data Augmentation Using Conditional Generative Models
• University project presentation (PDF). Explains the gist of the paper.
• In collaboration with Federico Vergallo

Prerequisites

• Python 3.6
• Tensorflow-gpu 1.14

Get data

bash resources/get_train_data.sh (~ 4.7 GB)
bash resources/get_test_data.sh (~1.8 GB)

Custom Training

$ python3 code/train.py
    --batch_size 128
    --num_classes 2
    --lr_g 0.0002
    --lr_d 0.005
    --model_name None
    --truncated False
    --rand_seed 42

Testing

$ python3 code/test.py
    --batch_size 128 -- default the number of images to generate
    --num_classes 2
    --model_name -- no default, download model and place in 'models/checkpoints'
    --truncated False
    --rand_seed 42

Pre-trained model

https://drive.google.com/file/d/1w7DMeCobR-GtRCgphfGIdZUo0Gwx5aDH/view?usp=sharing (~750mb)

sample run for 120 epochs

Architecture:

TO DO:

• explore time-of-day interpolation
• requirements.txt file

References

• Original attribute GAN paper AttGAN: Facial Attribute Editing by Only Changing What You Want