ccnet-jittor

This is a Jittor implementation of CCNet for semantic segmentation.

Quick Start

Environments

You should install Jittor under the guidance of Jittor official.

Training

To train with default batch-size 16 setting, you only need to specify the desired log directory, and run:

mpirun -np 4 python train.py --log_dir <log directory>

Testing

To test a model, you can specify the model checkpoint directory and the model's backbone, and run:

python test.py --ckpt_dir <log directory> --model_backbone <van or resnet>

Inference

To visualize a result, you can specify the picture's directory and the model, and run:

python visualize.py --ckpt_dir <log directory> --model_backbone <van or resnet> --pic_dir <picture directory> --save_dir <save directory>

Pretrained Checkpoints

Pretrained Backbone

Backbone	Checkpoints
ResNet101-C	TsinghuaCloud
VAN-base	TsinghuaCloud

CCNet Results

Settings	mIoU	Checkpoints
ResNet101-C, batch-size 16	41.72 (42.80 wi multiscale)	TsinghuaCloud
ResNet101-C, batch-size 8	38.85	TsinghuaCloud
ResNet101-C, batch-size 8, dilated	38.72	TsinghuaCloud
ResNet101-C, batch-size 8, neighborhood	36.77	TsinghuaCloud
VAN-base, batch-size 16	36.82	TsinghuaCloud