This repository demonstrates source codes and the dataset used in the paper "Disentangled Representation of Data Distributions in Scatterplots (IEEE InfoVis 2019 Short, PDF)".
We strongly recommend using Anaconda Python (>=3.6). Important dependencies are listed below.
torch==1.0.0
torchvision==0.2.2
- Clone this repository.
- Download the scatterplot images: https://www.kaggle.com/jaeminjo/scatterplotimages
- Create a subdirectory named
allunder theimagesdirectory, unzip the image archive, and put the images in the subdirectory. The source codes will look upimages/all/*.pngfor training images by default.
To train a β-VAE on the training images, run python train.py. This will generate a Torch model in the root directory (bvae.pt by default) as well as training logs (in a subdirectorytraining_logs/ by default). Note that the model and logs will be overwritten if you rerun the code. We also provide a pre-trained model, bvae_pretrained.pt.
usage: train.py [-h] [--data path] [--eta L] [--beta B] [--latents N]
[--batch-size N] [--img-size N] [--epochs N] [--no-cuda]
[--log-interval N] [--save-interval T] [--log-path path]
[--save-path path] [--resume path]
Train a beta-VAE using scatterplot images
optional arguments:
-h, --help show this help message and exit
--data path path to the scatterplot dataset (default: images)
--eta L learning rate for Adam (default: 1e-3)
--beta B the beta coefficient (default: 4)
--latents N number of latents (default: 32)
--batch-size N input batch size for training (default: 128)
--img-size N input image size (default: 64)
--epochs N number of epochs to train (default: 100)
--no-cuda disable CUDA training
--log-interval N how many batches to wait before logging training status
--save-interval T how many batches to wait before saving latent traversal
--log-path path output directory for training logs
--save-path path output path for the model
--resume path continue learning (from path)
To traverse the latent space, run python traverse.py <path_to_model>, e.g., python traverse.py bvae_pretrained.pt. This will perform latent traversals using images in the directory test_images as an input. The results will be saved in a separate directory (traverse_result by default). To reproduce Figure 1 in the paper, open figure.html on a Web browser after running the code.
Important Notes: Since the training procedure is stochastic, the order of latent dimensions from the largest KL divergence to the smallest will change if you rerun the training procedure. To include only the top N latent dimensions with the largest KL divergences in visualization, you need to provide the KL divergence of each latent dimension obtained from the training procedure. Open training_logs/log.txt and go to the last line. Copy the KL divergences (it is an array of as many floating values as the number of latent dimensions you used) and paste it into traverse.py, corr.py, and figure.html (search for 'TODO' in the codes to find the exact line). If you are using the pre-trained model, you can skip this.
usage: traverse.py [-h] [--input path] [--latents N] [--features N]
[--img-size N] [--partitions N] [--range N]
[--output output_path]
model_path
Traverse the latent space of beta-VAE using test images as an input
positional arguments:
model_path path to the trained model
optional arguments:
-h, --help show this help message and exit
--input path path to the input images
--latents N number of latents (default: 32)
--features N number of features to traverse (default: 8)
--img-size N input img size (default: 64)
--partitions N number of latent space partitions (default: 11)
--range N the maximum value for latent traversals (default: 2.5)
--output output_path path to the output directory
Run python corr.py <path_to_model>, e.g., python corr.py bvae_pretrained.pt to measure Pearson's ρ between human-perceived distances and predicted distances. By default, the script will train a neural network with a single hidden layer (with 32 neurons) to approximate the distance. All the latent dimensions will be used, but if you want to use the top N latent dimensions with the largest KL divergences, use the --bvae-features N flag. If you built your own model, you must provide the KL divergences of the latent dimensions. See Important Notes above. If you want to include scagnostics measures in the input, use the --use-scag flag.
Here are some results:
| Name | MSE Loss (Validation) | Pearson's ρ | Flag |
|---|---|---|---|
| Scagnostics | 0.124 | 0.644 | --use-scag --bvae-features 0 |
| β-VAE | 0.087 | 0.710 | Empty |
| β-VAE+Scagnostics | 0.070 | 0.750 | --use-scag |
usage: corr.py [-h] [--latents N] [--bvae-features N] [--img-size N]
[--hidden N] [--no-cuda] [--use-scag] [--folds N] [--lr L]
[--epochs N]
model_path
Compute the correlation between predicted distances and human-perceived
distances
positional arguments:
model_path path to the trained model
optional arguments:
-h, --help show this help message and exit
--latents N number of latents (default: 32)
--bvae-features N number of features to use (default: 32)
--img-size N input img size (default: 64)
--hidden N size of a hidden layer (default: 32)
--no-cuda disables CUDA training
--use-scag use scagnostics features?
--folds N number of folds
--lr L learning rate for Adam (default: 5e-3)
--epochs N number of epochs to train (default: 10000)
study_chi16contains the result, such as the human-perceived distances between scatterplots, from a previous study, "Towards Understanding Human Similarity Perception in the Analysis of Large Sets of Scatterplots (ACM CHI 16)" by Pandey et al. The original repository is https://github.com/nyuvis/scatter-plot-similarity.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent