Multiband GAN

Table of Contents

• About
• Getting started
• Experiments
• Reproducing results
• Web Application
• Credits

About

Multiband GAN is an adaptation of unsupervised method for continual learning of geneartive models called multiband training, introduced in Multiband VAE, to the Generative Adversarial Networks architecture. To stabilize the training, instead of standard GAN architecture, we use WGAN with gradient penalty. We are able to outperform other known methods on the most popular continual learning benchmarks.

Getting started

Prerequisites

To install this repository, you should have Python installed. This repository was tested under Linux and Windows 10.

Installation

1. Clone repository and navigate to project's directory

$ git clone https://github.com/bartooo/multiband_gan.git
$ cd multiband_gan/

2. Create new virtual environment

$ python -m venv gan_venv

3. Activate created virtual environment:

Linux:

$ source gan_venv/bin/activate

Windows:

$ gan_venv\Scripts\activate

4. Install requirements from requirements.txt file

$ pip install -r requirements.txt

Experiments

Multiband VAE

CL Benchmark	Num. of tasks	FID $\downarrow$	Precision $\uparrow$	Recall $\uparrow$
Split-MNIST Class Incremental	5	23	92	98
MNIST Dirichlet $\alpha=1$	10	30	92	97
Split-FashionMNIST Class Incremental	5	56	65	72
FashionMNIST Dirichlet $\alpha=1$	10	77	58	69
Split-Omniglot Class Incremental	5	12	98	96
Split-Omniglot Class Incremental	20	24	95	91
Omniglot Dirichlet $\alpha=1$	20	24	96	91
FashionMNIST $\rightarrow$ MNIST Class Incremental	10	49	68	70
MNIST $\rightarrow$ FashionMNIST Class Incremental	10	49	70	70
Split-CelebA Class Incremental	5	95	28.5	23.2
CelebA Dirichlet $\alpha=1$	10	93	33	22
CelebA Dirichlet $\alpha=100$	10	89	36.2	28

Multiband GAN

CL Benchmark	Num. of tasks	FID $\downarrow$	Precision $\uparrow$	Recall $\uparrow$
Split-MNIST Class Incremental	5	9	99	99
MNIST Dirichlet $\alpha=1$	10	20	98	98
Split-FashionMNIST Class Incremental	5	34	92	92
FashionMNIST Dirichlet $\alpha=1$	10	58	88	83
Split-Omniglot Class Incremental	5	1	99	98
Split-Omniglot Class Incremental	20	3	96	95
Omniglot Dirichlet $\alpha=1$	20	4	96	93
FashionMNIST $\rightarrow$ MNIST Class Incremental	10	31	89	87
MNIST $\rightarrow$ FashionMNIST Class Incremental	10	26	93	91
Split-CelebA Class Incremental	5	39	81	79
CelebA Dirichlet $\alpha=1$	10	60	67	67
CelebA Dirichlet $\alpha=100$	10	37	80	83
Split-CIFAR10 Class Incremental	5	43	92	79
Split-CIFAR100 Class Incremental	20	144	77	52

Reproducing results

1. Split-MNIST Class Incremental 5 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name CI_5 --dataset MNIST --gpuid 0 --seed 42 --num_batches 5 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 32 --g_n_features 32 --global_warmup 5 --only_generations

2. MNIST Dirichlet $\alpha=1$ 10 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name Dirichlet_alpha_1_10 --dataset MNIST --gpuid 0 --seed 42 --num_batches 10 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 32 --g_n_features 32 --dirichlet 1 --global_warmup 5 --only_generations

3. Split-FashionMNIST Class Incremental 5 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name CI_5 --dataset FashionMNIST --gpuid 0 --seed 42 --num_batches 5 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 32 --g_n_features 32 --global_warmup 5 --only_generations

4. FashionMNIST Dirichlet $\alpha=1$ 10 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name Dirichlet_alpha_1_10 --dataset FashionMNIST --gpuid 0 --seed 42 --num_batches 10 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 32 --g_n_features 32 --dirichlet 1 --global_warmup 5 --only_generations

5. Split-Omniglot Class Incremental 5 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name CI_5 --dataset Omniglot --gpuid 0 --seed 42 --num_batches 5 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 32 --g_n_features 32 --global_warmup 5 --only_generations

6. Split-Omniglot Class Incremental 20 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name CI_20 --dataset Omniglot --gpuid 0 --seed 42 --num_batches 20 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 32 --g_n_features 32 --global_warmup 5 --only_generations

7. Omniglot Dirichlet $\alpha=1$ 20 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name Dirichlet_alpha_1_20 --dataset Omniglot --gpuid 0 --seed 42 --num_batches 20 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 32 --g_n_features 32 --global_warmup 5 --dirichlet 1 --only_generations

8. FashionMNIST $\rightarrow$ MNIST Class Incremental 10 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name CI_10 --dataset DoubleMNIST --gpuid 0 --seed 42 --num_batches 10 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 32 --g_n_features 32 --global_warmup 5 --only_generations

9. MNIST $\rightarrow$ FashionMNIST Class Incremental 10 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name CI_10_reverse --dataset DoubleMNIST --gpuid 0 --seed 42 --num_batches 10 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 32 --g_n_features 32 --global_warmup 5 --reverse --only_generations

10. Split-CelebA Class Incremental 5 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name CI_5 --dataset CelebA --gpuid 0 --seed 42 --num_batches 5 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 50 --g_n_features 50 --global_warmup 5 --only_generations

11. CelebA Dirichlet $\alpha=1$ 10 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name Dirichlet_alpha_1_10 --dataset CelebA --gpuid 0 --seed 42 --num_batches 10 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 50 --g_n_features 50 --global_warmup 5 --dirichlet 1 --only_generations

12. CelebA Dirichlet $\alpha=100$ 10 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name Dirichlet_alpha_100_10 --dataset CelebA --gpuid 0 --seed 42 --num_batches 10 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 50 --g_n_features 50 --global_warmup 5 --dirichlet 100 --only_generations

13. Split-CIFAR10 Class Incremental 5 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name CI_5 --dataset CIFAR10 --gpuid 0 --seed 42 --num_batches 5 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 128 --g_n_features 128 --global_warmup 5 --only_generations

14. Split-CIFAR100 Class Incremental 20 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name CI_20 --dataset CIFAR100 --gpuid 0 --seed 42 --num_batches 20 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 128 --g_n_features 128 --global_warmup 5 --only_generations

15. CERN Class Incremental 5 tasks

python3 main.py --wandb_project MultibandGAN --experiment_name Class-Incremental_5_tasks --dataset CERN --gpuid 0 --seed 42 --num_batches 5 --latent_dim 100 --batch_size 64 --score_on_val --num_local_epochs 120 --num_global_epochs 200 --local_dis_lr 0.0002 --local_gen_lr 0.0002 --local_scheduler_rate 0.99 --lambda_gp 10 --num_gen_images 64 --local_b1 0.0 --local_b2 0.9 --global_gen_lr 0.001 --optim_noise_lr 0.1 --global_scheduler_rate 0.99 --limit_previous 0.5 --d_n_features 32 --g_n_features 32 --global_warmup 5 --only_generations

Application

Installation and running

NOTE: Gradio package works from Python 3.7 version.

1. Navigate to app's directory

$ cd app/

2. Create new virtual environment

$ python -m venv app_venv

3. Activate created virtual environment:

Linux:

$ source app_venv/bin/activate

Windows:

$ app_venv\Scripts\activate

4. Install requirements from requirements.txt file

$ pip install -r requirements.txt

Then, to run the application just run a demo.py file

python demo.py

Inference from random noise

Under the "Inference from random noise" tab, we can visualize the generations of the model trained according to the specific CL scenario and on a specific dataset. The generator takes as input a random noise vector.

Inference from provided noise

The "Inference from provided noise" tab is similar to the previously presented tab, except that now we can control the values of the input vector. The user provides the value of each element of that vector in the matrix. s a generation result, we now get the generation itself and the visualization of the translator’s latent space.

Credits

Repository based on Continual-Learning-Benchmark and Multiband VAE.