Known issues

Image is named as if it’s supported but has a non-supported format

The platform checks if a dataset fulfills our requirements when the dataset is uploaded to the platform.

If an image is saved in a non-supported format but still got a file-ending that is supported, e.g. .jpg, the dataset will pass the check.

Example: At the time of saving the user changed the extension (webpjpg) but that didn’t change the underlying format (webp) itself.

Then when a model uses the dataset, the training will crash, and the platform will show an error message since the file isn’t supported despite its name. For example, this:

Block configuration
[bound method TFInvalidArgumentErrorHandler.message of <potkaista.errors.handlers.TFInvalidArgumentErrorHandler object at 0x7ff2e500a7c0>]

Displaying npy files in the Datasets view

NumPy array data is visualized differently depending on dimensions.

The first dimension is always samples. If it has two dimensions, for example (100,2) we only write out the first value. If it has more dimensions, for example 3 (100,10,10) or 4 (100,10,10,3) we try to visualize it as an image (where width x height is 10x10) in grayscale or color.

If the data contains more than 4 dimensions we will fall back to only show the first value.

Was this page helpful?