This repository provides an implementation of Adversarially Learned Inference using a deeplearning library, Keras2.
- Python==3.5.*, 3.6.*
- Keras==2.0.*
- numpy==1.12.*, 1.13.*
- Pillow==4.*
All dependent libraries can be installed with pip command.
Note: We only tested our scripts using Keras with tensorflow backend.
You can run our training scripts with the following command.
$ python train.py [--epochs] [--batch_size] [--lr] [--beta_1] [--beta_2] [--snap_freq] [--result]
--epochs: training epochs.--batch_size: batch size during the training phase.--lr: learning rate (we use Adam as the optimizer).--beta_1: beta_1 parameter of Adam.--beta_2: beta_2 parameter of Adam.--snap_freq: save generaters' weights at each (snap_freq) epochs.--result: a path to the directory where training results (generater's weights and generated images) are to be saved.
ex) $ python train.py --epochs 500 --batch_size 100
All the default values of the above arguments and hyper parameters of the network architecture are basically the same values which are used in the original paper.
We used Titan X (pascal architecture) X 4 for all the experiments.
- 50,000 training images.
- 10,000 test images are used to generate or reconstruct images.
- All the images' shapes are unified into (h, w, c) = (32, 32, 3).
- The number of classes is 10 (5,000 images for each class).
epochs875(generate)
epochs875(reconstruct)
epochs3800(generate)
epochs3800(reconstruct)
epochs6000(generate)
epochs6000(reconstruct)
