Home

Awesome

deeplabv3p_gluon

DeepLab v3+ in MXNet Gluon

Note

git clone https://github.com/dmlc/gluon-cv
cd gluon-cv/scripts/datasets
python pascal_voc.py

Models

My porting on Pascal VOC validation:

ModelEvalOS (w/ or w/o inference tricks)mIoU (%)
xception_coco_voc_trainaug (TF release)16 (w/o) <br> 8 (w/)82.20 <br> 83.58
xception_coco_voc_trainaug (MXNet porting)16 (w/o) <br> 8 (w/o)79.19<br>81.82
xception_coco_voc_trainaug (MXNet finetune TrainOS=16)16 (w/o) <br> 8 (w/o)82.75<br>82.56
xception_coco_voc_trainaug (MXNet finetune TrainOS=8)16 (w/o) <br> 8 (w/o)82.02<br>83.14
xception_voc_trainaug <br> ImageNet pretrained only, without MSCOCO pretrained16 (w/o) <br> 8 (w/o)77.06<br>76.44

AWS Runtime & Cost

Measured with fixing batch stats (use_global_stats=True), just for reference.

InstanceGPUsPricingTrain OSSpeedTrain on train_augEval on valTime per epochCost per epoch
p2.8xlargeK80x87.20$/h16<br>81.5s/b16<br>3.4s/b1617.0min<br>37.5min3.5min<br>10min<br>(BUGS: gpus do not use sufficiently during eval)20.5min<br>47.5min$2.5<br>$5.7
p3.8xlargeV100x412.24$/h16<br>80.5s/b16<br>3.0s/b125.5min<br>44.5min0.7min<br>1.3min6.2min<br>45.8min$1.3<br>$9.3

Memo

Acknowledge

This repository is a part of MXNet summer code hosted by AWS, TuSimple and Jiangmen. Specifically, I would like to thank Hang Zhang (@AWS) and Hengchen Dai (@TuSimple) for kind suggestions on tuning and implementation. Plus, I would like to thank AWS for providing generous credits for tuning the computationally intensive models.