• Added a new v2_transform() method to replace torch.cat + nn.Conv2d combinations with CatConv2d, an all-in-1 fused cuda kernel combining Concat and Conv2d.

• Inference on TitanV speedup from 70 fps to 99 fps

• CatConv2d Installation

cd CatConv2d/
python setup.py install

(Please note that backward path for CatConv2d hasn't been implemented)

• Implementaion based on Harmonic DenseNet: A low memory traffic network (ICCV 2019)
• Refer to Pytorch-HarDNet for more information about the backbone model
• This repo was forked from meetshah1995/pytorch-semseg

• Simple U-shaped encoder-decoder structure
• Conv3x3/Conv1x1 only (including the first layer)
• No self-attention layer or Pyramid Pooling

• ** Speed tested in 1536x768 instead of full resolution.

• CamVid
• Pascal VOC
• ADE20K
• MIT Scene Parsing Benchmark
• Cityscapes
• NYUDv2
• Sun-RGBD

• pytorch >=0.4.0
• torchvision ==0.2.0
• scipy
• tqdm
• tensorboardX

Setup config file

Please see the usage section in meetshah1995/pytorch-semseg

To train the model :

python train.py [-h] [--config [CONFIG]]
--config                Configuration file to use (default: hardnet.yml)

To validate the model :

usage: validate.py [-h] [--config [CONFIG]] [--model_path [MODEL_PATH]] [--save_image]
                       [--eval_flip] [--measure_time]
  --config              Config file to be used
  --model_path          Path to the saved model
  --eval_flip           Enable evaluation with flipped image | False by default
  --measure_time        Enable evaluation with time (fps) measurement | True by default
  --save_image          Enable writing result images to out_rgb (pred label blended images) and out_predID

• Cityscapes pretrained weights: Download
  (Val mIoU: 77.7, Test mIoU: 75.9)
• Cityscapes pretrained with color jitter augmentation: Download
  (Val mIoU: 77.4, Test mIoU: 76.0)
• HarDNet-Petite weights pretrained by ImageNet:
  included in weights/hardnet_petite_base.pth