Configurations

This document provides an overview of the config.yaml file used for configuring the project and explains the available command-line options for overriding configurations.

Overview of config.yaml

The config.yaml file organizes the project configuration into three main sections: DATA, directories, and training.

Note

The YAML file is not required for running the project with default settings. The project is designed to run seamlessly without a config.yaml file by using pre-defined default configuration values.

Default configurations

Data Configuration

The DATA section specifies the dataset and its properties:

dataset:
  dataset_path: "erikpinhasov/landcover_dataset"
  batch_size: 16
  num_workers: 8
  do_reduce_labels: False
  pin_memory: True

Details:

  • dataset_path: Path to the dataset. In this case, it references a Hugging Face dataset.

  • batch_size: Number of samples processed per training batch.

  • num_workers: Number of worker threads for data loading.

  • do_reduce_labels: Whether to apply label reduction.

  • pin_memory: If True, enables faster GPU transfer by pinning memory.

  • id2label: Maps class indices to their corresponding labels.

Project Configuration

The directories section defines directory paths for saving models, logs, and results:

directories:
  models: models
  pretrained: models/pretrained_models
  logs: models/logs
  checkpoints: models/logs/checkpoints
  results: results

Details:

  • models: Directory to save trained models.

  • pretrained: Directory for storing pre-trained models.

  • logs: Directory for logs.

  • checkpoints: Directory for storing training checkpoints.

  • results: Directory for evaluation results.

Training Configuration

The training section provides options for training parameters:

training:
  model_name: b0
  max_epochs: 50
  learning_rate: 2e-5
  weight_decay: 1e-3
  ignore_index: 0
  weighting_strategy: "raw"
  alpha: 0.7
  id2label:
    0: "background"
    1: "bareland"
    2: "rangeland"
    3: "developed space"
    4: "road"
    5: "tree"
    6: "water"
    7: "agriculture land"
    8: "buildings"

Details:

  • model_name: Name of the SegFormer variant (e.g., b0).

  • max_epochs: Maximum number of epochs for training.

  • learning_rate: Learning rate for the optimizer.

  • weight_decay: Strength of L2 regularization applied to the optimizer.

  • ignore_index: Label index to ignore during training (int or None).

  • weighting_strategy: Method for normalizing class weights ("none", "balanced", "max", "sum", "raw").

  • alpha: Cross-Entroy loss alpha weight parameter, beta = 1 - alpha.

  • id2label: Mapping from class indices to class labels.

Callbacks Configuration

The callbacks section defines parameters for managing callbacks during training:

callbacks:
  early_stop_patience: 5
  early_stop_monitor: "val_loss"
  early_stop_mode: "min"
  save_model_monitor: "val_mean_iou"
  save_model_mode: "max"

Details:

  • early_stop_patience: Number of epochs to wait for improvement before stopping training.

  • early_stop_monitor: Metric to monitor for early stopping (e.g., "val_loss").

  • early_stop_mode: Direction of improvement ("min" for decreasing metrics, "max" for increasing metrics).

  • save_model_monitor: Metric to track for saving the best model (e.g., "val_mean_iou").

  • save_model_mode: Direction of improvement ("min" for decreasing metrics, "max" for increasing metrics).

Command-Line Options

Command-line options allow overriding config.yaml settings or default values. For example:

  • Specify a custom configuration file:

    ucs --config path/to/config.yaml train

  • Override specific parameters without a configuration file:

    ucs train --batch_size 32 --learning_rate 0.001

For more details, refer to the Usage Guide.