Creating your datasets

Datasets manage images and labels assigned to those images. Training a model requires a training dataset (trains the model) and a test dataset (evaluates the trained model). To train a model with Defect Detection App, you must create a training dataset, and optionally you can create a test dataset. If you only create a training dataset, Defect Detection Station App, trains the model by splitting the training dataset to create a test dataset. If you want control over the images that Defect Detection App uses to evaluate a model, you can create a test dataset. For example, you might want to use a set of benchmark images to compare evaluation results between different models thjat you train.

You create a dataset by using the Defect Detection App Console. During dataset creation you upload the images that you want to use to train the model. You can choose to automatically classify the images as normal or anomaly. This saves you from having to classify the images in the image gallery. After you create the dataset, you label the images according to the type of model that you want to create (image classification or image segmentation). You can add more images to the dataset, as you need them. For example, if the training evaluation results are poor, you can improve the model by adding more images to the dataset. Note that you can't delete images from a dataset.

The following procedure shows how to create a training dataset and optionally a test dataset.

To create your datasets

1. Sign in to the Defect Detection App Console.

2. In the top navigation pane, choose Projects.

3. On the projects page, choose the project for the dataset.

4. The Training dataset tab displays the dataset details. Initially, the dataset contains no images.

5. In the Images panel, choose Add images.

6. In the Add images to dataset page, choose a classification method for the images that you upload. You can classify the images as normal, anomaly, no classification, or use a file prefix to classify the image. If you choose to use a file prefix, images without the matching prefixes are added to the dataset, but you will need to classify them. To use a file prefix, do the following:

   1. Enter the file prefix for normal images in Normal prefix.

   2. Enter the file prefix for anomalous images in Anomaly prefix.

7. Choose Upload files and choose the images that you want to use. Alternatively, drag and drop the image files to the page. If you used the Station App to capture images, the images are in the folder you noted in step 6 of Capturing images (Station App). You can upload up to 30 images at a time.

8. Choose Save to finish uploading the images.

9. Repeat the Add images step until you have added all the images you need to create the dataset. To guide you, note the training requirements for the training dataset (and test requirements, if you're creating a test dataset).

10. (Optional) Create a test dataset by doing the following:

    1. On the project details page, choose the Test dataset tab.

    2. Choose Create test dataset.

    3. Repeat steps 6 - 9 to add images to your test dataset.

11. Next step: Annotating dataset images.