PIDNet

Thanks to the PIDNet !:kissing_heart::kissing_heart::kissing_heart:

PIDNet Paper | PIDNet Code

💛What's new?

In the official version, we have made the following changes:

1. 💥Adjusted the number of GPUs for training to one, and handled some Bugs!!
2. 💥Add <tools/FLOPS_demo.py>, which is easy for getting Flops and params!!
3. 💥Add some useful demos in <./demo>!!!
4. 💥Made a cheat sheet to facilitate a quick start!!!

💙How to train PIDNet for a customed datasets.

Here, we use pidnet-small to make an example. You need to change these settings one by one:

1. You need to prepare cityscapes-style image data in folder as following:

   • data/cityscapes/custom_datasets/
     • gtFine # grayscale images, e.g. containing only 0 and 255
     • images # RBG images or some else

2. Make a [.lst] file for the trainloader to read images as following:

   • data/cityscapes/list/cityscapes/
     • train.lst # like 'custom_datasets/images/1.jpg custom_datasets/gtFine/1.png', one for images, one for gtFine
     • val.lst # if you don't have val data, or you can use a copy of train.lst and rename it to val.lst)

3. Change dataset setting as following, and make sure len(self.class_weights) = len(self.label_mapping):

   • datasets/cityscapes.py
     • self.label_mapping
     • self.class_weights

4. Set necessary configs in [.yaml] file, which is in:

   • configs/cityscapes/pidnet_small_cityscapes.yaml
     • set DATASET.TEST_SET, DATASET.TRAIN_SET and DATASET.NUM_CLASSES
     • set TRAIN.IMAGE_SIZE, TRAIN.BASE_SIZE, TEST.IMAGE_SIZE, TEST.BASE_SIZE
     • set TRAIN.BATCH_SIZE_PER_GPU, TRAIN.END_EPOCH if you like

5. Train

  python tools/train.py --cfg configs/cityscapes/pidnet_small_cityscapes.yaml

6. Your output will be saved in:
   • output/cityscapes/pidnet_small_cityscapes/

💚How to inference?

1. Inference

By the default setting, you need to put your images in for inference.

len(color_map in <tools/inference.py>) >= the class number of your dataset is okay.

  cd tools
  python inference.py --c [your_dataset_class_num] --p ../output/cityscapes/pidnet_small_cityscapes/checkpoint.pth.tar

2. Where's the output?

The output will be saved to <samples/output/>