Model Training

OceanEye uses the ONNX (Open Neural Network Exchange) format, which is a popular cross-platform format for sharing AI models.

It is reccomended to use a CUDA-capable GPU. CUDA can be downloaded here. While training on a CPU is supported, it may result in significantly longer training times.

Model training requires additional dependencies, which are automatically installed during the initial training process with OceanEye. An internet connection is required to install dependencies.

It is recommended to have at least 200 annotations per class before initiating training. For optimal results, having over 1,000 annotations per class is strongly encouraged.

Model Training Settings

Setting	Description
Model Save Location	Filepath to the outputted model

Advanced Settings

Setting	Description
Base Model	Specifies the filename of a supported Ultralytics model to be used as the starting point for training.
Training Time	Defines the maximum allowable duration for training. Set to 0 for unlimited training time.
Max Epochs	Indicates the total number of epochs for training, where each epoch represents a complete iteration over the dataset.
Patience	Determines the number of consecutive epochs without improvement in validation metrics before terminating the training process early.

Model training requires Python. An isolated Python environment is automatically created within the .oceaneye directory located in your home folder. Approximately 5 GB of additional dependencies will be installed during this process. Depending on your system's performance, training may take several hours or even multiple days. It is recommended to allow the process to run uninterrupted.