super-gradients/src/super_gradients/recipes/cityscapes_stdc_seg50.yaml

#  STDC segmentation training example with Cityscapes dataset.
#  Reproduction and refinement of paper: Rethinking BiSeNet For Real-time Semantic Segmentation.
#
#  Usage STDC1-Seg50:
#      python -m torch.distributed.launch --nproc_per_node=2 train_from_recipe.py --config-name=cityscapes_stdc_seg50 checkpoint_params.external_checkpoint_path=<stdc1-backbone-pretrained-path>
#  Usage STDC2-Seg50:
#      python -m torch.distributed.launch --nproc_per_node=2 train_from_recipe.py --config-name=cityscapes_stdc_seg50 checkpoint_params.external_checkpoint_path=<stdc1-backbone-pretrained-path> architecture=stdc2_seg
#
#
#  Validation mIoU - Cityscapes, training time:
#      STDC1-Seg50:    input-size: [512, 1024]     mIoU: 75.11     2 X RTX A5000, 20 H
#      STDC2-Seg50:    input-size: [512, 1024]     mIoU: 76.44     2 X RTX A5000, 23 H
#
#  Official git repo:
#      https://github.com/MichaelFan01/STDC-Seg
#  Paper:
#      https://arxiv.org/abs/2104.13188
#
#  Pretrained checkpoints:
#      Backbones- downloaded from the author's official repo.
#       https://deci-pretrained-models.s3.amazonaws.com/stdc_backbones/stdc1_imagenet_pretrained.pth
#       https://deci-pretrained-models.s3.amazonaws.com/stdc_backbones/stdc2_imagenet_pretrained.pth
#
#      Logs, tensorboards and network checkpoints:
#       STDC1-Seg50: https://deci-pretrained-models.s3.amazonaws.com/cityscapes_stdc1_seg50_dice_edge/
#       STDC2-Seg50: https://deci-pretrained-models.s3.amazonaws.com/cityscapes_stdc2_seg50_dice_edge/
#
#  Learning rate and batch size parameters, using 2 RTX A5000 with DDP:
#      STDC1-Seg50:    input-size: [512, 1024]     initial_lr: 0.01    batch-size: 16 * 2gpus = 32
#      STDC2-Seg50:    input-size: [512, 1024]     initial_lr: 0.01    batch-size: 16 * 2gpus = 32
#
#  Comments:
#      * Pretrained backbones were used.

defaults:
  - cityscapes_stdc_base
  - _self_

dataset_params:
  _convert_: all
  batch_size: 16
  val_batch_size: 16
  crop_size: [1024, 512]
  eval_scale: 0.5
  random_scales: [ 0.125, 1.5 ]

training_hyperparams:
  loss:
    dice_ce_edge_loss:
      num_classes: 19
      ignore_index: 19
      weights: [ 1., 0.6, 0.4, 1. ]
      dice_ce_weights: [ 1., 1. ]
      ce_edge_weights: [ .5, .5 ]
      edge_kernel: 3

experiment_name: ${architecture}50_cityscapes