Features

• Graphs in ClearML - Display graphs for any Darknet compatible networks.

  • For classifiers:

    • Accuracy (acc);
    • Average loss (avg_loss);
    • learning rate (lr).

  • For detectors:

    • Mean average precision 50% (mAP@0.50);
    • Precision;
    • Recall;
    • F1-score;
    • Average loss (avg_loss);
    • learning rate (lr).

• Logs in ClearML - Displays logs from Darknet in ClearML.

• Saving weights in ClearML - Ability to save weights in ClearML with flexible adjustment of parameters.

  Available options:

  • Darknet.is_save_weights - Ability to disable saving weights (for experiments).
  • Darknet.save_weights_from_n_iterations - Ability to save weights from the start of N iterations. For example, N=10000.
  • Darknet.save_weights_every_n_iterations - Ability to save weights every N iterations. For example, N=5000.

• Additional function of data splitting - Allows you to divide the data set into selections.

  📚 Click to see example

  from clearml import Dataset
  from clearml_darknet.utils import split_dataset
  
  
  dataset_path = Dataset.get(dataset_name='dataset-example', dataset_project='Tests/darknet').get_local_copy()
  
  train, valid = split_dataset(
    dataset_path=dataset_path,
    ratio=0.7,
    shuffle=True
  )

Requirements

• Python (3.6)+
• ClearML (1.4.1)
• Darknet

Installation

Before running, be sure to clone the Darknet repository and compile for the appropriate operating system.

$ pip install clearml-darknet-py

Examples

• An example of a detector training script - train_detector.py
• An example of a classificator training script - train_classifier.py
• Detector optimization script example - optimize_detector.py

For examples and use cases, check the examples folder.

License

Clearml Darknet is MIT licensed as listed in the LICENSE file.