PyTorch DeepLabV3

PyTorch implementation of DeepLabV3

Results

• Training: PASCAL VOC 2012 trainaug set
• Evaluation: PASCAL VOC 2012 val set

Backbone	Output stride	mIoU	Configuration
ResNet101	16	77.8	configs/resnet101.yaml
MobileNetV2	16	70.1	configs/mobilenetV2.yaml

Requirements

• Python ≥ 3.6
• Install libraries: pip install -r requirements.txt

Usage

Data preparation

cd datasets/voc/

# Download standard PASCAL VOC 2012
wget http://pjreddie.com/media/files/VOCtrainval_11-May-2012.tar
tar xf VOCtrainval_11-May-2012.tar

# The original PASCAL VOC 2012 only labels 1,464 training images. Following the paper, we add additional
# annotations from Semantic Boundaries Dataset to augment the training set to 10,582 images.
wget http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/semantic_contours/benchmark.tgz
tar zxf benchmark.tgz

# Generate CSV files that contain the paths to the images and annotations.
python prepare.py --voc VOCdevkit/VOC2012/ --sbd benchmark_RELEASE/

Configuration

We use YAML for configuration management. See configs/*.yaml for examples. You can modify the settings as needed.

Training

python train.py --cfg <CONFIG_FILE> --logdir <LOG_DIRECTORY>

# For example, to train a model with ResNet101 backbone:
python train.py --cfg configs/resnet101.yaml --logdir runs/resnet101/exp0/

To visualize training progress using TensorBoard:

tensorboard --logdir <LOG_DIRECTORY>

An interrupted training can be resumed by:

# Run train.py with --resume to restore the latest saved checkpoint file in the log directory.
python train.py --cfg <CONFIG_FILE> --logdir <LOG_DIRECTORY> --resume

Evaluation

python eval.py --cfg <CONFIG_FILE> --pth <LOG_DIRECTORY>/best.pth --csv datasets/voc/val.csv