Frédéric TOST
Generated images
This is a Python/Tensorflow 2.0 implementation of the Adversarial Latent AutoEncoders. See reference below:
- Stanislav Pidhorskyi, Donald A. Adjeroh, and Gianfranco Doretto. Adversarial Latent Autoencoders. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR), 2020. [to appear]
preprint on arXiv: 2004.04467
MNIST dataset is used as a toy example. The Generator and E encoder are using Conv2D and Conv2DTranspose instead of a MLP (Multi-Layer Perceptron used in the paper). This gives better results but a longer training.
The objective is to show how to easily implement the ALAE using the MNIST dataset and convolutional networks. Finding the hyperparameters such as learning rate of each optimizer is the most fastidious task.
Additional features
- The use of Tensorflow 2.0 HParams Dashboard features allows to keep a trace of each run using different hyperparameters.
- 3 losses time history
| Loss | Time history |
|---|---|
| Generator |  |
| Discriminator |  |
| Latent |  |
-
the latent loss was splitted into Reconstruction loss and Kullback Leibler (KL) loss. KL loss is not used in the original paper, it seems to accelerate the convergence of the Generator/Discriminator.
K_RECONST_KL = 1.0 # To use full reconstruction (alae_tf2.py)
To run the demo, you will need to have installed Tensorflow 2.0.0 or more recent (2.1.0, 2.2.0).
Run the demo
python alae_tf2.py
| Path | Description |
|---|---|
| alae_tf2.py | Configure the hyperparameters and run the demo. |
| alae_tf2_helper.py | Train the neural network. alae_helper is the class that define the losses and the train step function. |
| alae_tf2_models.py | Models used in the demo, Encoders F & E, Generator and Discriminator (4 classes). |
| utils.py | Useful functions to process and plot images (2 functions). |
- Frédéric TOST — Developer TOST Corp.