Input and output data

Specifying input and output types

You can configure the input and output data types for the functions evaluate, train and createState. To do so, go to the "Configure" tab on the model's page, click the gear wheel settings icon. Under "Input/Output", click "Edit Inputs/Outputs". Then, select which function to edit parameters for. Evaluate has both input data and output data, while train and createState only have input data. Each operation can take multiple parameters. Give each parameter a name, a set of allowed data types and a shape. Data types and shapes will soon be explained.

You may need to modify the code of the model so that it is able to use the new parameters.

If you are running operations using the API you will also have to update your application so that it provides the correct input parameters and is able to interpret the new output parameters. If you attempt to run an operation using the API but are providing incorrect input data types, the operation will not be started and will instead give you an error.

Data types

The available data types are:

Numbers come in various forms and sizes. f32 and f64 are 32-bit and 64-bit floats respectively, which means numbers that can have a decimal point. Since f64 has more bits, it is more precise and can hold larger numbers. u8, u16, u32 and u64 are 8-bit, 16-bit, 32-bit and 64-bit unsigned integers respectively, which means that they cannot hold negative values. i8, i16, i32 and i64 are 8-bit, 16-bit, 32-bit and 64-bit signed integers respetively, which means that they can hold negative values. For integers and unsigned integers, a larger bit size means that the value can hold larger numbers. u8 can hold values between 0 and 255, while u64 can hold values between 0 and roughly nine quintillion.

The dictionary data type is a bit special - it is a data type which contains key/value pairs, where keys are text and values are data. Each key/value pair can have a unique data type and also a unique shape. This can for example be used to represent a 2D point - you would have two keys "x" and "y", both of type "number" and shape [].

If you have a data type that doesn't really fit into any of the above mentioned types, you can always choose binary and manage the serialization and deserialization yourself.

Data shape

Input and output parameters are actually lists of data, and possibly nested lists of data. That is, a list containing another list that contains data.

The shape specifies how many elements each list should have, and how deep the nesting goes. For example, if the data type is set to image, the shape "[1]" would mean that the model expects to receive a list containing one image, which equates to a single image in total. The shape "[2]" would mean that the model expects to receive a pair of images. The shape "[-1]" means that any number of images can be provided. The shape "[2, 3]" would mean that the model expects to receive two arrays that each contain three images, for a total of six images. The shape "[2, -1, 3]" would mean that the model expects two arrays that contain any number of arrays that each contain three elements. The empty shape, "[]", means that the data is just a single element, not a list. Input and output parameters cannot have the shape "[]". To specify a single element you should instead specify a list that contains a single element ("[1]"). Dictionary entries can have the shape "[]", as noted in the example above.

Try it out

Below is the parameter editor that you can use to define the input/output parameter types of your model, and the data editor which you can use to provide data for a parameter. Try changing the parameter types, and you can then try to provide data for that type in the editor below. These UI elements are not linked to any model - try changing things to see what happens!