Ebm_compositionality
| Project Name | Stars | Downloads | Repos Using This | Packages Using This | Most Recent Commit | Total Releases | Latest Release | Open Issues | License | Language |
|---|---|---|---|---|---|---|---|---|---|---|
| Gans Awesome Applications | 4,667 | a month ago | 17 | |||||||
| Curated list of awesome GAN applications and demo | ||||||||||
| 3d Pointcloud | 1,753 | 9 days ago | 2 | Python | ||||||
| Papers and Datasets about Point Cloud. | ||||||||||
| Awesome Question Answering | 658 | 7 months ago | ||||||||
| Resources, datasets, papers on Question Answering | ||||||||||
| Vipassana For Hackers | 582 | 4 months ago | 3 | cc-by-sa-4.0 | TeX | |||||
| A document version of my "Vipassana for Hackers" talk | ||||||||||
| Recipenlg | 505 | 2 years ago | Jupyter Notebook | |||||||
| Set of scripts and notebooks used to produce results visible in RecipeNLG paper | ||||||||||
| Conditional Pixelcnn Decoder | 470 | 5 years ago | 3 | Python | ||||||
| Tensorflow implementation of Gated Conditional Pixel Convolutional Neural Network | ||||||||||
| Question Generation Paper List | 429 | 2 years ago | ||||||||
| A summary of must-read papers for Neural Question Generation (NQG) | ||||||||||
| Human Video Generation | 345 | 4 months ago | 1 | |||||||
| Human Video Generation Paper List | ||||||||||
| Awesome Few Shot Image Generation | 295 | 4 months ago | ||||||||
| A curated list of papers, code and resources pertaining to few-shot image generation. | ||||||||||
| Handwriting Generation | 289 | 6 years ago | 9 | mit | Python | |||||
| Implementation of handwriting generation with use of recurrent neural networks in tensorflow. Based on Alex Graves paper (https://arxiv.org/abs/1308.0850). | ||||||||||
Code for Compositional Visual Generation with Energy Based Models A pytorch codebase for compositionality can be found here.
Install Prerequisites
Please install the required python packages by running the command below:
pip install -r requirements.txt
Download Datasets
We run experiments on Mujoco Scenes and CelebA dataset. To generate data used in the Mujoco Scenes dataset, look in the image_comb directory (you will need to appropriately modify the path) and run the corresponding files inside. For example to generate the continual learning dataset, you can use the command:
python image_comb/cube_continual.py
Feel free to reach out to us for pre-generated Mujoco Scenes Datasets
You can download the CelebA dataset here
Training
Models are trained using the following command:
python train.py --dataset=<dataset> --exp=<exp_name> --cclass --step_lr=100.0 --swish_act --num_steps=60 --num_gpus=<gpu_num>
Evaluation
The files ebm_sandbox.py and celeba_combine.py contains evaluation functions used to reproduce results in the paper. Different models can be set in the celeba_combine.py file, and different tasks evaluated using the --task flag in ebm_sandbox.py. You can use the command below to generate compositions of young, female, smiling and wavy hair faces:
python celeba_combine.py
High Resolution CelebA Generation
High resolution images in CelebA are composed using the training method here. Code for composing and training models can be found here as well as pretrained models.
Cubes Dataset
The dataset used for 3D cube experiments can be found at:
https://www.dropbox.com/sh/202zhctt6rac0lw/AACAYhk6K6_FPYrremx9A1D_a?dl=0
