Awesome

DCGAN-pytorch (LSGAN)

The code is modified from https://github.com/pytorch/examples/tree/master/dcgan.

Prerequisites

• Python 3.6
• GPU Memory >= 2G

Getting started

Installation

• Install Pytorch from http://pytorch.org/
• Install Torchvision from the source

git clone https://github.com/pytorch/vision
cd vision
python setup.py install

Because pytorch and torchvision are ongoing projects.

Here we noted that our code is tested based on Pytorch 1.0.0 and Torchvision 0.2.0.

mkdir model
mkdir visual
mkdir result

Train

python train.py --dataset market --dataroot /home/zzd/market1501/pytorch/train_all  --withoutE --name baseline-lsgan8x8-encode --lsgan --gpu_ids 3

--name the name of the output model

--lsgan mean using MSELoss(L2Loss)

--gpu_ids select which gpu to run

--batchSize default 64

by default, batch norm is used. Conv layers have no bias.

--instance to use instance norm. Conv layers also learn bias.

--withoutE to remove Encoder Network. The code will run as the basic LSGAN.

Now I set step learning rate schedule. The learning rate drop 0.1 at 40th epoch.

Test(Generate Images)

python test.py  --name baseline  --batchsize 16  --which_epoch 24