ffjord - GitLite

KO | EN

gitlite — search

#python #hacktoberfest #ai #llm #golang #machine-learning #linux #docker #nodejs #awesome #kubernetes #vue #awesome-list #mcp

ffjord

(★ 669)

code for "FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Models".

File Explorer

Download Latest Version (.zip)

- github_flow.gif
- __init__.py
- bsds300.py
- gas.py
- hepmass.py
- LICENSE.txt
- miniboone.py
- power.py
- approx_error_1d.py
- plot_bottleneck_losses.py
- plot_compare_multiscale.py
- plot_flows.py
- plot_nfe_vs_dim_vae.py
- plot_sn_losses.py
- scrap_log.py
- viz_cnf.py
- viz_high_fidelity_toy.py
- viz_multiscale.py
- viz_toy.py
- github.png
- maple_leaf.jpg
- - __init__.py
  - cnf.py
  - container.py
  - coupling.py
  - elemwise.py
  - glow.py
  - norm_flows.py
  - normalization.py
  - odefunc.py
  - resnet.py
  - squeeze.py
- custom_optimizers.py
- datasets.py
- multiscale_parallel.py
- odenvp.py
- priors.py
- spectral_norm.py
- toy_data.py
- utils.py
- visualize_flow.py
- - __init__.py
  - CNFVAE.py
  - flows.py
  - layers.py
  - VAE.py
- - __init__.py
  - loss.py
  - training.py
- - __init__.py
  - distributions.py
  - load_data.py
  - log_likelihood.py
  - plotting.py
  - visual_evaluation.py
- LICENSE
.gitignore
LICENSE
README.md
train_cnf.py
train_discrete_tabular.py
train_discrete_toy.py
train_img2d.py
train_misc.py
train_tabular.py
train_toy.py
train_vae_flow.py

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Free-form Jacobian of Reversible Dynamics (FFJORD)

Code for reproducing the experiments in the paper:

> Will Grathwohl*, Ricky T. Q. Chen*, Jesse Bettencourt, Ilya Sutskever, David Duvenaud. "FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Models." _International Conference on Learning Representations_ (2019).
> [[arxiv]](https://arxiv.org/abs/1810.01367) [[bibtex]](http://www.cs.toronto.edu/~rtqichen/bibtex/ffjord.bib)

## Prerequisites

Install `torchdiffeq` from https://github.com/rtqichen/torchdiffeq.

## Usage

Different scripts are provided for different datasets. To see all options, use the `-h` flag.

Toy 2d:
```
python train_toy.py --data 8gaussians --dims 64-64-64 --layer_type concatsquash --save experiment1
```

Tabular datasets from [MAF](https://github.com/gpapamak/maf):
```
python train_tabular.py --data miniboone --nhidden 2 --hdim_factor 20 --num_blocks 1 --nonlinearity softplus --batch_size 1000 --lr 1e-3
```

MNIST/CIFAR10:
```
python train_cnf.py --data mnist --dims 64,64,64 --strides 1,1,1,1 --num_blocks 2 --layer_type concat --multiscale True --rademacher True
```

VAE Experiments (based on [Sylvester VAE](https://github.com/riannevdberg/sylvester-flows)):
```
python train_vae_flow.py --dataset mnist --flow cnf_rank --rank 64 --dims 1024-1024 --num_blocks 2
```

Glow / Real NVP experiments are run using `train_discrete_toy.py` and `train_discrete_tabular.py`.

## Datasets

### Tabular (UCI + BSDS300)
Follow instructions from https://github.com/gpapamak/maf and place them in `data/`.

### VAE datasets
Follow instructions from https://github.com/riannevdberg/sylvester-flows and place them in `data/`.

## Bespoke Flows

Here's a fun script that you can use to create your own 2D flow from an image!
```
python train_img2d.py --img imgs/github.png --save github_flow
```

<p align="center">
<img align="middle" src="./assets/github_flow.gif" width="400" height="400" />
</p>