DeepLab - Pytorch

Kevin Wang - COMS4998: A Deep Learning Project

Repo is based on https://github.com/VainF/DeepLabV3Plus-Pytorch

1. Install Requirements

pip install -r requirements.txt

2. Train the model on Cityscapes

2.1 Download and Extract Dataset

Download the CityScapes Dataset. You will need a free account to do so. Download "gtFine" and "leftImg8bit" datasets.

Extract two zip files gtFine_trainvaltest.zip, leftImg8bit_trainvaltest.zip and extract it to the following folder structure:

/datasets
    /data
        /cityscapes
            /gtFine
            /leftImg8bit

2.2 Start Visdom Server (Visualize Training)

# First start visdom server on port 28333
visdom -port 28333

2.1 Run Training Command on Cityscapes Dataset

py main.py --model deeplabv3plus_mobilenet --dataset cityscapes --enable_vis --vis_port 28333 --gpu_id 0  --lr 0.1  --crop_size 768 --batch_size 2 --output_stride 16 --data_root ./datasets/data/cityscapes 

3. Make your Prediction

Our examples are in test_images and results are stored in test_results

Single image:

py predict.py --input /test_images/1325466082.jpg  --dataset cityscapes --model deeplabv3plus_mobilenet --ckpt checkpoints/best_deeplabv3plus_mobilenet_cityscapes_os16.pth --save_val_results_to test_results

Image folder:

py predict.py --input /test_images --dataset cityscapes --model deeplabv3plus_mobilenet --ckpt checkpoints/best_deeplabv3plus_mobilenet_cityscapes_os16.pth --save_val_results_to test_results