Skip to content
This repository has been archived by the owner on Apr 19, 2024. It is now read-only.
/ MONet-pytorch Public archive

Burgess et al. "MONet: Unsupervised Scene Decomposition and Representation"

License

Notifications You must be signed in to change notification settings

baudm/MONet-pytorch

Repository files navigation

MONet in PyTorch

We provide a PyTorch implementation of MONet.

This project is built on top of the CycleGAN/pix2pix code written by Jun-Yan Zhu and Taesung Park, and supported by Tongzhou Wang.

Note: The implementation is developed and tested on Python 3.7 and PyTorch 1.1.

Implementation details

Decoder Negative Log-Likelihood (NLL) loss

Decoder NLL loss

where I is the number of pixels in the image, and K is the number of mixture components. The inner term of the loss function is implemented using torch.logsumexp(). Each pixel of the decoder output is assumed to be iid Gaussian, where sigma (the "component scale") is a fixed scalar (See Section B.1 of the Supplementary Material).

Test Results

CLEVR 64x64 @ 160 epochs

The first three rows correspond to the Attention network outputs (masks), raw Component VAE (CVAE) outputs, and the masked CVAE outputs, respectively. Each column corresponds to one of the K mixture components.

For the fourth row, the first image is the ground truth while the second one is the composite image created by the pixel-wise addition of the K component images (third row).

CLEVR 1

CLEVR 2

Prerequisites

  • Linux or macOS (not tested)
  • Python 3.7
  • CPU or NVIDIA GPU + CUDA 10 + CuDNN

Getting Started

Installation

  • Clone this repo:
git clone https://github.com/baudm/MONet-pytorch.git
cd MONet-pytorch
  • Install [PyTorch](http://pytorch.org and) 1.1+ and other dependencies (e.g., torchvision, visdom and dominate).
    • For pip users, please type the command pip install -r requirements.txt.
    • For Conda users, we provide a installation script ./scripts/conda_deps.sh. Alternatively, you can create a new Conda environment using conda env create -f environment.yml.
    • For Docker users, we provide the pre-built Docker image and Dockerfile. Please refer to our Docker page.

MONet train/test

  • Download a MONet dataset (e.g. CLEVR):
wget -cN https://dl.fbaipublicfiles.com/clevr/CLEVR_v1.0.zip
  • To view training results and loss plots, run python -m visdom.server and click the URL http://localhost:8097.
  • Train a model:
python train.py --dataroot ./datasets/CLEVR_v1.0 --name clevr_monet --model monet

To see more intermediate results, check out ./checkpoints/clevr_monet/web/index.html.

To generate a montage of the model outputs like the ones shown above:

./scripts/test_monet.sh
./scripts/generate_monet_montage.sh

Apply a pre-trained model

  • Download pretrained weights for CLEVR 64x64:
./scripts/download_monet_model.sh clevr

About

Burgess et al. "MONet: Unsupervised Scene Decomposition and Representation"

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published