Skip to content

robertopossidente/Distributed-DCGAN

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Deep Convolution Generative Adversarial Networks

This repository is a distributed data parallel version of dcgan.pytoch examples

The example implements the paper Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks

The implementation is very close to the Torch implementation dcgan.torch

After every 100 training iterations, the files real_samples*.png and fake_samples*.png are written to disk with the samples from the generative model.

Downloading Dataset

First download and extract the desired dataset. In this example we'll download and use CIFAR-10 dataset, which can be obtained using the commands below:

mkdir cifar10 && cd cifar10
wget --no-check-certificate https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
tar -xvf cifar-10-python.tar.gz
cd ..

This will download and extract cifar-10 dataset to ./cifar10/

Installing and Running

  1. Build the docker image from this repository using the command below:
docker build -t dist_dcgan .
  1. Let's start 2 terminals to test the network using data distributed parallelism. On first terminal run:
docker run --env OMP_NUM_THREADS=1 --rm --network=host -v=$(pwd):/root dist_dcgan:latest python -m torch.distributed.launch --nproc_per_node=1 --nnodes=2 --node_rank=0 --master_addr="172.17.0.1" --master_port=1234 dist_dcgan.py --dataset cifar10 --dataroot ./cifar10

And in the second terminal, run:

docker run --env OMP_NUM_THREADS=1 --rm --network=host -v=$(pwd):/root dist_dcgan:latest python -m torch.distributed.launch --nproc_per_node=1 --nnodes=2 --node_rank=1 --master_addr="172.17.0.1" --master_port=1234 dist_dcgan.py --dataset cifar10 --dataroot ./cifar10
  • The -v parameter is mapping the current directory to docker's /root directory, where data and the scripts resides.
  • The --nproc_per_node is the number of processes to start in each node.
  • The --nnodes parameter defines the number total of nodes (world_size).
  • The --node_rank parameter defines the rank for this particular node. Each node must have an unique node_rank.
  • The --master_addr and -master_port parameters is the address and port to the master node (rank 0), respectively. Check if docker's IP is visible otherwise the processing will be stuck and will never starts.
  • The --dataset and --dataroot parameters indicates the dataset and where it is located, respectively.
  • Use --gpus=all to docker command and --cuda to dist_dcgan.py to enable the use of GPUs. For multiple gpus in a single node, it is recommended to set --nproc_per_node=1 (one process per gpu) and each for each process, use the --local_rank parameter to assign it a a specific GPU (or using CUDA_VISIBLE_DEVICES environment variable).

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 98.9%
  • Dockerfile 1.1%