DeepLabV3Plus for Beginners

Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation

simple version(w/o inplace_abn) is here

Support & Requirements

pytorch >= 1.1.0
tensorboardX
CUDA >= 10.0
SyncBN support(inplace_abn)

Install Inplace_abn

$ pip install git+https://github.com/mapillary/inplace_abn.git@v1.0.10

Train

git clone & change DIR

$ git clone https://github.com/J911/DeepLabV3Plus-for-Beginners
$ cd DeepLabV3Plus-for-Beginners

run 🙌🙌

$ python -m torch.distributed.launch --nproc_per_node ${num of GPUs} train.py --data /data/CITYSCAPES --batch-size 16 --epoch 200 --logdir ./logs/exp1/ --save ./saved_model/exp1/

Evaluate

$ python -m torch.distributed.launch --nproc_per_node 1 evaluate.py --data /data/CITYSCAPES --weight ./saved_model/exp1/epoch200.pth --num-classes 19

Dataset

This Repository uses Cityscapes Dataset.

CITYSCAPES
|-- leftImg8bit
|   |-- test 
|   |-- train
|   `-- val
`-- gtFine
    |-- test 
    |-- train
    `-- val

Result

TBD

Thanks to

@speedinghzl - Gain a lot of Insight 🙇🏻‍♂️
@mapillary - using inplace_abn