FAQ¶

In an industrial setting, with a controlled environment, identical lighting, and singular subject matter, a few dozen images per class might be enough. If the data is obtained in real-world conditions, you will need hundreds or more annotated input samples per class for a usable model.

Types supported: The following image types are supported: JPG, JPEG, PNG, 8-bit TIF and TIFF, BMP, GIF, including a 4^th transparency channel. Higher bit-depth grayscale images (14-bit and 16-bit) are automatically converted to 8-bit during import using right 8-bit shifting. Note that 14-bit grayscale images must be MSB-aligned (data bits aligned to the Most Significant Bit side) for correct conversion. EXIF orientation metadata is supported.

During dataset creation, to accurately evaluate your model, you can split your data into two datasets to have data for training and validation. The training set is the largest part of your dataset that you reserve for training your model. The validation set is a separate set of your samples that you will use during training to track how your model performs on unseen data. Based on the validation dataset, you can make the training stop if the model stops improving. This is called early stopping and can be configured during training setup.

During training setup, when you use the auto setup option, all data in your project is split into two datasets: 80% for training and 20% for validation.

You can create a third split, a test set, by manually creating one dataset containing samples and annotations to test on. This test set can be used in the test center to test your model on.

Alternatively, you can reuse your validation set as the test set in the test center. We don't use the validation set for hyperparameter tuning; therefore, leakage is minimal. Only through the early stopping.

Key metrics presented by our platform and how they change as training progresses can help you understand if you are overfitting or if you are unnecessarily training for too long. If your model performs much better on the training set than on the test set, then you are likely overfitting. It means your model does not generalize well from your training data to unseen data. For example, if the accuracy of your model on the training set is 99%, but on the validation set it is only 55%.

It is especially helpful to check the Training loss curve vs the Validation loss curve:

• If your Training loss curve does not decrease, you are underfitting. The model is not learning.
• If your Training loss curve is much lower than your Validation loss curve, you are overfitting.
• If the Validation loss curve decreases for a while and then increases, the minimum is where you should stop training (with early stopping, it's done automatically).

To prevent overfitting, try one or a combination of the following:

• Get more training data, especially of rare classes.
• Turn on the Early stopping toggle (default).
• Set Save strategy to Overwrite best (default).
• Add the Dropout rate.

This is an overview of some common training parameters. Each model will have its own specific parameters.

Use GPU for training: Enabling the GPU will drastically speed up the training.

Epochs: The maximum number of times the training data will be processed by the training algorithm (Min. 1).

Batch size: The number of data samples in a single neural network weights update. When the value is too large, it is possible that there will not be enough GPU memory for a single network update. In general, a higher batch size is better, as long as your GPU has sufficient memory.

Freeze backbone: Whether or not the weights of the backbone are frozen during the training. With a smaller amount of data, it is better to freeze the backbone.

Early stopping: Stops the training earlier when the model is no longer improving.

Early stopping patience: The number of epochs with no model improvement after which the training will be stopped (Min. 1).

Learning rate: The magnitude of the changes to the neural network weights during training. The learning rate will decrease during the training if Learning rate decay is set to a value greater than 0 (Min. 0).

Save strategy: When Overwrite best is selected, the model that is currently performing best on the validation set is saved, but it will be overwritten if a better model comes along. This method guarantees you still have a usable model if the training should fail.

Transfer learning: You can reuse the weights from another model for the model you are about to train. This gives the training a head start and leads to a better model. The classes in both models should be the same or similar. The model parameters will be set to the ones used in the selected model to ensure compatibility. It is possible but not recommended to alter these model parameters.

Image size: The network expects a fixed input resolution for all inputs, so all images will be scaled to a square image with the width and height (in pixels) defined by this parameter (Min. 64).

Models are the result of training sessions. When a model is trained on a Robovision AI platform, this training is shown in the training center of this platform. In contrast, when you import an inference setup from a local storage or from another Robovision AI platform, the training was done in another platform. As a result, imported models aren't displayed on the training details page and the Training center section of the project details page because trainings—not models—are shown there. However, you will still be able to select imported models for transfer learning, to set up a test, and more.

To use the Prediction tool for data labeling, you need a well-trained model capable of generating reliable annotations.

The Prediction tool can be used to:

• Pre-label new data to speed up manual annotation.

  We recommend enabling predictive labeling before starting labeling so that all labelers receive the same suggestions, improving consistency across annotations.

• Re-annotate existing data with an improved or optimized model.

• Compare model-generated predictions with manual labels for evaluation and quality control.

On the menu, go to System > About. In the General section, you can view the version of your Robovision AI instance.

On the menu, go to System > About. For support information, view the Support section. For technical help, start by downloading logs from the Logs section and verifying your platform version in the General section. Once these steps are completed, get in touch with our support team and provide the gathered information.