Usage¶

This document provides guidance on how to use the Urban Climate Segmentation Model repository for training, evaluating, and managing outputs efficiently.

Default Configurations¶

The project is designed to run seamlessly even if config.yaml does not exist. When no configuration file is provided, it automatically uses default settings. If a config.yaml file exists, it will be loaded automatically. Refer to the Configurations Guide for details.

Training the Model¶

Use the ucs train command to start training. Examples:

Default Training:
```
ucs train
```
With Custom Configurations:
```
ucs train --config path/to/config.yaml
```
Command-Line Overrides:
```
ucs train --batch_size 32 --lr 0.001
```

Training Outputs¶

During training, the following outputs are generated:

Checkpoints: Saved in models/logs/checkpoints/. This includes the best model checkpoints based on validation performance.
Pretrained Models: The pretrained BERT-based transformer model from Hugging Face is saved in models/pretrained_models/.
Lightning Logs: Training logs for TensorBoard are stored in models/logs/lightning_logs/. These logs include metrics and other details for visualization.

To view all available training options, run:

ucs train -h

Evaluating the Model¶

Evaluate a specific version of the model and save results, including a confusion matrix, to the results/ directory.

Running Evaluation¶

Default Evaluation:
```
ucs evaluate
```

Evaluate Specific Version:

ucs evaluate -v <version>

Example:

ucs evaluate -v 5

Evaluation Outputs¶

Confusion Matrix: Saved in the results/ directory as:
```
results/version_<version>_confusion_matrix.png
```

Folder Structure and Outputs¶

During training and evaluation, outputs are organized as follows:

models/: Root folder for all model-related outputs.
- models/logs/:
  - models/logs/checkpoints/: Contains the best model checkpoints.
  - models/logs/lightning_logs/: Training logs for TensorBoard.
- models/pretrained_models/:
  - Contains pretrained models used for initializing training.
results/:
- Contains evaluation results, including confusion matrices saved as:
  - results/version_<version>_confusion_matrix.png

For more details, refer to the Configurations Guide.