Launch: Train Larger Models on Roboflow

With Roboflow Train, you can train computer vision models in the cloud. Your models can then be deployed to your hardware with Workflows and Roboflow Inference.

We are excited to announce that customers on our Basic and Growth plans can now train larger models on Roboflow Train. Larger model sizes are available for Object Detection, Instance Segmentation, and Keypoint models. The new model sizes are only available for the Roboflow 3.0 model architecture.

The larger models we now support – Medium, Large and Extra Large – will run slower than the default Fast model architecture, but will be more accurate.

Larger model sizes are ideal for when you are working on projects where accuracy is essential.

In this guide, we will walk through how to train a model using one of the larger model sizes.

Train a Large Model with Roboflow

To get started, open a Project in your Roboflow Workspace, then generate a dataset version. A version is a snapshot of your dataset frozen in time. You can train a model from a dataset version.

Once you have a dataset version ready, click “Custom Train”:

A window will pop up asking you to select an architecture. This release adds larger model training to the Roboflow 3.0 architecture. Select “Roboflow 3.0” as the model architecture to see the larger model options.

This window shows three new sizes: Medium, Large and Extra Large.

The most accurate model size for object detection, segmentation,and keypoint detection we now offer is “Extra Large”.

Select a model size, then click “Continue”.

You will then be asked to choose a checkpoint from which to train:

When you train your first model of a given size, we recommend training from the Microsoft COCO checkpoint. After you train your first model with a given size, you will be able to train from the previous model using the “Train from Previous Checkpoint” option. This is ideal when you already have a model that works well that you want to improve.

Click “Start Training” to start the model training job.

Your dataset will be prepared for training, then you will receive an estimate for how long we think the model training job will take. The larger models we have launched will take longer to train than the Fast model architecture.

You will then be able to monitor your training job as the model trains. Live training graphs will appear that show the performance of your model.

Conclusion

You can now train Medium, Large, and Extra Large models when you choose Roboflow 3.0 as your training architecture on Roboflow Train. These model sizes are available for object detection, instance segmentation, and keypoint detection models on Roboflow.

To learn more about Roboflow Train, refer to our model training documentation.