Skip to content

DMML-Geneva/ksgan

Repository files navigation

The official implementation of: "Kolmogorov–Smirnov GAN"

by Maciej Falkiewicz and Naoya Takeishi and Alexandros Kalousis

The arXiv preprint can be found under this link.

Project structure

The experimental pipeline

The project is governed with Data Version Control (DVC) with the pipeline defined in dvc.yaml and parameterized with a global config file: experiments/configs/global.yaml. There are two types of datasets: synthetic datasets and real-world datasets which have independent pipelines, both defined in dvc.yaml.

Synthetic datasets pipeline

flowchart TD
	node1["evaluate@{model_id}-{dataset_name}-{simulation_budget}-{model_seed}"]
	node2["generate@{dataset_name}"]
	node3["preprocess@{dataset_name}"]
	node4["train@{model_id}-{dataset_name}-{simulation_budget}-{model_seed}"]
	node5["visualize-data@{dataset_name}"]
	node6["visualize-evaluation@{model_id}-{dataset_name}-{simulation_budget}-{model_seed}"]
	node7["visualize-training@{model_id}-{dataset_name}-{simulation_budget}-{model_seed}"]
	node1-->node6
	node2-->node3
	node3-->node1
	node3-->node4
	node3-->node5
	node4-->node1
	node5-->node1
	node4-->node7
Loading

The possible values of the variables are:

  • {model_id}: "model1" (GAN), "model2" (WGAN), "model3" (KSGAN)
  • {dataset_name}: "swissroll", "circles", "rings", "moons", "8gaussians", "pinwheel", "2spirals", "checkerboard"
  • {simulation_budget}: "512", "1024", "2048", "16384", "65536"
  • {model_seed}: "0", "1", "2", "3", "4"

Real-world datasets pipeline

flowchart TD
	node1["evaluate-real@{model_id}-{dataset_name}-{model_seed}"]
	node2["preprocess-real@{dataset_name}"]
	node3["train-real@{model_id}-{dataset_name}-{model_seed}"]
	node4["visualize-training-real@{model_id}-{dataset_name}-{model_seed}"]
	node2-->node1
	node2-->node3
	node3-->node1
	node3-->node4
Loading

The possible values of the variables are:

  • {model_id}: "model1" (GAN), "model2" (WGAN), "model3" (KSGAN)
  • {dataset_name}: "mnist", "cifar10"
  • {model_seed}: "0", "1", "2", "3", "4"

Execution on a slurm orchestrated cluster

Attention: the scripts assume that preprocess@{dataset_name} (preprocess-real@{dataset_name}) stages have been already executed!

Please mind that this way of execution bypasses DVC, and thus requires commiting the changes in order to control versions.

In CPU_PARTITIONS and GPU_PARTITIONS environmental variables you should specify the available CPU and GPU partitions.

Source code

Environment

All the python dependencies are listed in requirements.txt file.

Data

The DVC cache for the project can be downloaded from here.

About

Kolmogorov–Smirnov GAN

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published