4. Train model

Once you have labeled (some of) your images, you can start with training a first model on your dataset.

To get started quickly, you can use our preconfigured notebooks. However, if you want to train your model locally, or use another cloud infrastructure like Amazon Sagemaker, look at how to export your dataset and how to push your trained model to our platform to use model-assisted labeling.

To get access to our notebooks, go to the model tab and click on the green "plus" icon. Here you can select which architecture you would like to train. For object detection projects, we currently support YOLOv5 and YOLOv8, while for instance segmentation projects we only support YOLOv8 for now.

If you click on one of the proposed architectures, a Colab notebook will open.

To get started, fill in your API key and Project Name. You can change the default parameters like im_size, n_epochs, batch_size, and model if you like, but it is a good starting point. Next, click "Runtime" => "Run all" to start the training process.

You can find your API key by clicking on your avatar next to your username (top right)

Once finished training, you should see a new version added to the model's tab. Here, you can compare your new version with previous versions. Also, you can start using this new version to accelerate your labeling using our model-assisted labeling tool.

Different model versions.

Train/validation split

Once you start a training session, your data is automatically split into an 80/20 split, meaning 80% of your images are assigned as training data, while the other 20% is assigned as validation data.

For some use cases, it might be beneficial to manually change some images from the training set to the validation set or visa versa. You can do this by going to the data tab and clicking on the drop-down list of each image.