sud0301 / 8gpu_test

cityscapes_training

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

8gpu_test

cityscapes_training

Dataset preparation

Cityscapes

Download the dataset from the Cityscapes dataset server(Link). Download the files named 'gtFine_trainvaltest.zip', 'leftImg8bit_trainvaltest.zip' and extract in ./data/cityscapes/

Create the environment

conda create -n 8gpu_test python=3.9
pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
pip install numpy

Training on Cityscapes Dataset on 8-GPUs

python train_full.py    --batch-size 40

Training on Cityscapes Dataset with 8 Jobs on 1 GPU

./train_8x1gpu.sh

Reduce batch-size, if GPU memory is not enough.

About

cityscapes_training


Languages

Language:Python 99.6%Language:Shell 0.4%