super-gradients/src/super_gradients/recipes/cityscapes_regseg48.yaml

#  RegSeg segmentation training example with Cityscapes dataset.
#  Reproduction of paper: Rethink Dilated Convolution for Real-time Semantic Segmentation.
#

# Instructions:
#   0. Make sure that the data is stored in dataset_params.dataset_dir or add "dataset_params.data_dir=<PATH-TO-DATASET>" at the end of the command below (feel free to check ReadMe)
#   1. Move to the project root (where you will find the ReadMe and src folder)
#   2. Run the command:
#       python src/super_gradients/examples/train_from_recipe_example/train_from_recipe.py --config-name=cityscapes_regseg48
#
#
#  Validation mIoU - Cityscapes, training time:
#      RegSeg48:    input-size: [1024, 2048]     mIoU: 78.15  using 4 GeForce RTX 2080 Ti with DDP, ~2 minutes / epoch
#
#  Official git repo:
#      https://github.com/RolandGao/RegSeg
#  Paper:
#      https://arxiv.org/pdf/2111.09957.pdf
#
#
#  Logs, tensorboards and network checkpoints:
#      s3://deci-pretrained-models/regseg48_cityscapes/
#
#
#  Learning rate and batch size parameters, using 4 GeForce RTX 2080 Ti with DDP:
#      RegSeg48:    input-size: [1024, 2048]     initial_lr: 0.02    batch-size: 4 * 4gpus = 16

defaults:
  - training_hyperparams: default_train_params
  - dataset_params: cityscapes_regseg48_dataset_params
  - checkpoint_params: default_checkpoint_params
  - _self_

train_dataloader: cityscapes_train
val_dataloader: cityscapes_val

cityscapes_ignored_label: 19    # convenience parameter since it is used in many places in the YAML

architecture: regseg48

arch_params:
  num_classes: 19
  strict_load: no_key_matching

load_checkpoint: False

ckpt_root_dir:

resume: False
training_hyperparams:
  sync_bn: True
  resume: ${resume}
  max_epochs: 800
  lr_mode: poly
  initial_lr: 0.02   # for effective batch_size=16
  lr_warmup_epochs: 0
  optimizer: SGD
  optimizer_params:
    momentum: 0.9
    weight_decay: 5e-4

  ema: True

  loss: cross_entropy
  criterion_params:
    ignore_index: ${cityscapes_ignored_label}

  train_metrics_list:
    - PixelAccuracy:
        ignore_label: ${cityscapes_ignored_label}
    - IoU:
        num_classes: 20
        ignore_index: ${cityscapes_ignored_label}

  valid_metrics_list:
    - PixelAccuracy:
        ignore_label: ${cityscapes_ignored_label}
    - IoU:
        num_classes: 20
        ignore_index: ${cityscapes_ignored_label}

  metric_to_watch: IoU
  greater_metric_to_watch_is_better: True

  _convert_: all

project_name: RegSeg
experiment_name: ${architecture}_cityscapes

multi_gpu: AUTO
num_gpus: 4


# THE FOLLOWING PARAMS ARE DIRECTLY USED BY HYDRA
hydra:
  run:
    # Set the output directory (i.e. where .hydra folder that logs all the input params will be generated)
    dir: ${hydra_output_dir:${ckpt_root_dir}, ${experiment_name}}