This repository implements general network for semantic segmentation.
You can train various networks like DeepLabV3+, PSPNet, UNet, etc., just by writing the config file.
You can run pretrained DeepLabv3+ converted from official tensorflow model.
Currently I checked that xception65_cityscapes_trainfine can be converted.
$ mkdir tf_model
$ cd tf_model
$ wget http://download.tensorflow.org/models/deeplabv3_cityscapes_train_2018_02_06.tar.gz
$ tar -xvf deeplabv3_cityscapes_train_2018_02_06.tar.gz
$ cd ../src
$ mkdir ../model
$ python convert.py ../tf_model/deeplabv3_cityscapes_train/model.ckpt 19 ../model/cityscapes_deeplab_v3_plus/model.pth
Then you can test the performance of trained network.
$ python eval.py
In order to train model, you have only to setup config file.
For example, write config file as below and save it as config/pascal_unet_res18_scse.yaml.
Net:
enc_type: 'resnet18'
dec_type: 'unet_scse'
num_filters: 8
pretrained: True
Data:
dataset: 'pascal'
preprocess: 'imagenet'
target_size: (256, 256)
Train:
max_epoch: 20
batch_size: 2
resume: False
start_epoch: 0
Loss:
weight:
size_average: True
batch_average: True
Optimizer:
mode: 'adam'
base_lr: 0.001
t_max: 10Then you can train this model by:
$ python train.py ../config/pascal_unet_res18_scse.yaml
- Cityscapes
- Pascal Voc
.
├── config
├── data
│ ├── cityscapes
│ │ ├── gtFine
│ │ └── leftImg8bit
│ └── pascal_voc_2012
│ └── VOCdevkit
│ └── VOC2012
│ ├── JPEGImages
│ ├── SegmentationClass
│ └── SegmentationClassAug
├── logs
├── model
└── src
├── dataset
├── logger
├── losses
│ ├── binary
│ └── multi
├── models
│ └── inplace_abn
└── utils
- OS: Ubuntu18.04
- python: 3.6.4
- pytorch: 0.4.1