Skip to main content

Importing Datasets

Importing a dataset also prepares the dataset for use with the APU. The FVS input in fp32 or binary npy format, or binary bin format – Datasets that have been converted via an inference utility from media databases of videos, images, or music.

To import a dataset:

  1. From the FVS page, click Select file. The file selector appears.

    Select File

  2. Choose your dataset vector file. Click Upload.

  3. In the Datasets section, click Browse. The Datasets window appears.

    Browse Dataset

  4. Click Add Dataset. The Create new dataset window appears.

    Create New Dataset

  5. Click the file you want and click OK. Additional dataset parameter fields appear.

  6. Enter the dataset parameters.

    Dataset Parameters Selection

  7. If you choose the IVF search type, you will need to indicate the following:

    • The number of nBits per dataset.

    • The number of nBits per cluster.

    • In order to split the dataset into clusters: select the number of centroids or according to the number of instances/boards.

      • If you select the centroids according to the number of instances/boards, the total centroids nBits per instance is 128K x 768. For example, if nBits = 768, then the number of centroids per instance is 128K. If nBits = 256, then the number of centroids per instance is 384K.

      • In case of multiple instances/boards, then the total number of centroids is the number of instances/boards x 128K x 768.

        Cluster according to a number of centroids:

        Centroids Clustering

        Cluster according to the number of instances/boards:

        Instance Clustering

  8. Click OK. The dataset is uploaded and trained.