How to change the resolution of a model training run in Roboflow
Sticking with the default input size usually yields the best balance of accuracy, speed, and stability. It is often better to:
- Tile images, if you are dealing with small objects
- Changing the resize strategy, instead of the size itself, especially if you are concerned about the square size
- Choosing a larger model size entirely
- Adding more representative training data to your dataset
You can change the resolution of your dataset training during the version creation process, specifically by changing the settings of the Resize step in the Preprocessing section.
If you do not specify or resize the default value for that model architecture will be used, which differs for each model architecture & model architecture size.
Note: Higher resolutions are not guaranteed to improve accuracy and may increase the latency of the model. Higher resolution trainings are also more susceptible to out-of-memory errors during the training process.