Fork of Hugging Face Diffusers by Petr Zahradník with additional features and fixes for 1D diffusion.
Courtesy of the master's thesis "Diffusion models for path planning". Will be published soon.
Features added to 1D U-Net:
- Coalesce U-Net block classes for consistency
- Allow tuning of many internal parameters (normalization, activation, padding, time embedding)
- Support for Dropout
- Support for LoRA convolution
- Simplify ResNet skip connections handling
Why? Upstream implementation follows closely Janner's Diffuser. Our fork diverges to simplify both the implementation and usage. The code is rewritten to follow the 2D U-Net which is much more advanced.
This fork is not published on pip. It must be installed from the source code:
pip install .Refer to the Diffusers documentation, this fork is mostly compatible.
The training script used in the thesis is in train.py.
License: Apache-2.0 by HuggingFace Team and Petr Zahradník