Probabilistic Dataset of Abstract Shapes
- This library provides a latent factor model that can be applied to any rendering function in a flexible way. It can sample inputs to the rendering function, and evaluate their likelihood. It can support multiple different distributions at the same time (for instance, in-distribution and out-of-distribution, or different environments).
- Additionally, it provides 'Dsprites++' as a rendering frontend, supporting colors, textures, and more.
- It provides sensible defaults for creating a meaningful dataset, ready generated and provided for download.
Possible applications for the dataset:
- Out-of-distribution detection
- Concept discovery
- Disentanglement
- Causal discovery
- Domain transfer, domain adaption, few-shot learning, etc.
- Generative modeling/density estimation
- [x] Distributions for the latent factor model
- [x] Latent factor model with sampling and likelihoods, multiple distributions
- [ ] Support for latent SCMs
- [ ] Simple parallel rendering via a .sample_parallel() method
- [x] Basic shapes rendering
- [x] Textures, perlin noise, other post-processing
- [ ] Arrangements of multiple foreground shapes in a specified
The script example.py shows how to define a model with a default distribution (in-distribution) and several other distributions that differ in various ways (out-of-distribon, OoD).
