Training models
If you've yet to create a model, head over to creating a model for help getting started. Otherwise, read on to learn how to train your models on Encord.
Training models is currently only supported on image sequences and videos. Support for images will be following soon! Please let us know at support@encord.com if this is something you're interested in!
Training a model
The first screen on the model training dashboard i.e. ‘Training settings’ displays the model type, the relevant ontology objects and allows users to define:
- Number of epochs
- Advanced settings:
- Processing unit: CPU or GPU
- Batch size
- Training weights: pre-trained weights to use; these vary according to the selected framework and model architecture
After making the appropriate choices, the second screen offers the option to select which data units to use for training. In the example above, we only have two image groups to select from. In this case, we have only selected the reviewed image group for training.
Once you're ready, press Train to start the process!
Monitor training progress
Once a training job is launched, you can monitor its progress from the notifications tab, accessed via the bell icon in the upper right navigation bar. While the model is training, you will be updated with the current epoch number and loss value.
Once the training has finished, a green check mark will appear along with an indication of how long ago the training job completed.
Congratulations! You've successfully trained a model, proceed to our section on Inference to learn how to apply your model to generate automated labels.
Model training logs
Once model training is finished, we can explore the training logs through the relevant option under the model tab. From here we can download the training log or visualise the loss curve in the UI.