mobilenetv3-segmentation

An unofficial implement of MobileNetV3 for semantic segmentation.

Requisites

• PyTorch 1.1

• Python 3.x

Usage

Train

• Single GPU training

python train.py --model mobilenetv3_small --dataset citys --lr 0.0001 --epochs 240

• Multi-GPU training

# for example, train mobilenetv3 with 4 GPUs:
export NGPUS=4
python -m torch.distributed.launch --nproc_per_node=$NGPUS train.py --model mobilenetv3_small --dataset citys --lr 0.0001 --epochs 240

Evaluation

• Single GPU training

python eval.py --model mobilenetv3_small --dataset citys

• Multi-GPU training

# for example, evaluate mobilenetv3 with 4 GPUs:
export NGPUS=4
python -m torch.distributed.launch --nproc_per_node=$NGPUS --model mobilenetv3_small --dataset citys

Result

• Cityscapes

where: lr=0.01, crop_size=768

Note: Params and Madds are got using torchscope. They are much larger than those reported in the paper.

To Do

•  improve performance

•  train and eval

•  test madds

References

• maskrcnn-benchmark