Train
Our managed computer vision training solution will give you a state of the art model hosted at an API endpoint customized for your dataset in no time.
Roboflow offers an AutoML product called Roboflow Train. It is the easiest way to train and deploy a state of the art object detection model on your custom dataset. It's literally one click -- we'll do the rest. When your model is done training, you'll receive access to a hosted inference API to interrogate your model for predictions via your programming language of choice (or a simple demo web app), a Tensorflow JS model you can embed in your web application, and an on-device inference server you can run on edge devices like the NVIDIA Jetson.
Roboflow Train credits are available in your workspace depending on which plan you have selected for the workspace. Training a model costs one credit and takes between 1 and 24 hours depending on the size of your dataset. Contact our sales team to upgrade your plan if you need more train credits.
- Contribute to the Roboflow Blog: Author a post, and share your work with Roboflow blog and newsletter readers.
- Apply for Research Credits: Additional training credits and increased account limits are available for research and education.
Choose image preprocessing and image augmentation settings, then generate a version of your dataset.
For Object Detection projects, there is a "Fast" model and an "Accurate" model available. For all other project types, there is only one training checkpoint option available.
- Accessing the Accurate model for training: Reach out to the sales team (businesses), or file a Community ("Apply for Research Credits") or Blog Contributor form (students, researchers, and hobbyists). NOTE: The Accurate model is available by default for customers on the Growth and Enterprise plans.
- The default image resize is 640x640. If you select a "Resize" preprocessing option, we will train a model whose native input size is similar to your training data's size in the generated version.
- To make your model faster, try exporting images in a smaller size.
The maximum size we recommend for Roboflow Train is 1280x1280 (1024x1024 for Semantic Segmentation projects). However, for the vast majority of use cases, 640x640 is best (512x512 for Semantic Segmentation projects).
If you have images larger than this, it is recommend you employ Tiling as a preprocessing step, and Resize your images to a square size between 640x640 and 1280x1280 -- larger images are quite slow to train and conduct inference against. So if you're not sure what to pick, we again recommend starting with 640x640.
The larger the dataset, and the larger your outputted (generated) images, the longer it will take for your dataset to train. We will email you when it's finished. In most cases, this should be under 24 hours.
You can choose to train your model from the Microsoft COCO dataset checkpoint, one of the over 10,000 trained models on Roboflow Universe, or from "Scratch."
Training from a Checkpoint means that you are employing Transfer Learning. Transfer Learning will initialize your model training from the model you have selected. This can help to reduce training time, and provide you with improved training scores.
Training from Scratch means that you are not employing Transfer Learning. This will initialize your model training with randomized initial values for the model weights.
If you are looking to benchmark or prototype a model with different preprocessing and augmentation settings, or different combinations of images, then it is best to train from the same, pre-trained Checkpoint (i.e training each version from the Microsoft COCO Checkpoint, the same Public Checkpoint, or the same "Previous" Checkpoint from one of your own projects).
Training Options for Object Detection Projects
Training Options for Classification Projects
Training Options for Instance Segmentation Projects
Training Options for Semantic Segmentation Projects
We've added the option to train from any trained model in Roboflow Universe that matches your project type (i.e Object Detection checkpoints are only available for Object Detection projects).
When training from checkpoints, if the model checkpoint you are training from is the Fast model, then the checkpoint will be available under the Fast model option. If the model checkpoint you are training from is the Accurate model, then the checkpoint you will have available is under the Accurate option.
"Previous Checkpoints" are model checkpoints from any project in a workspace that you are currently part of, and hold an
Admin or Creator role in.NOTE: When training from checkpoints, if the model checkpoint you are training from is the Fast model, then the checkpoint will be available under the Fast model option. If the model checkpoint you are training from is the Accurate model, then the checkpoint you will have available is under the Accurate option. Additionally, checkpoints are only available for Transfer Learning from the same project type.
- You can find the Model Type for a trained version on that project version's page
This project version was trained with Roboflow Train's Fast model
This option is recommended for those training datasets on "novel" or niche projects, where there are no existing trained models with similar subjects to your data (e.g, astrophotography).
We've made available a Model Library to help you more easily train a custom model architecture of your choice with your data.
When your model has finished training, you can see the metrics on the dataset Versions page including mean Average Precision (mAP), precision, recall, and more.
- You can save your training graph by right-clicking it and selecting "Save Image As."
After training, your model is ready to be used for inference and embedded in a custom application! See the Inference Documentation page for all the options.
Last modified 4mo ago