Choose a single data type for all features in your training dataset: INT8, INT16, or FLOAT32. For mixed-type datasets, choose the widest type present (for example, FLOAT32 if any value is a float).
Input data type specification allows you to optimize the preprocessing operation and reduce SRAM and NVM usage on the device, and inference time. The data of the same type must be used for prediction. You must select the correct data type, as incorrect selection of input data type may cause target metric degradation or total footprint increase.
You only need to choose one data type for the entire dataset. For example, if you have a dataset with three features and two of them are represented as INT16 while the third one is represented as FLOAT32, you should choose FLOAT32 as the data type for the entire dataset. The platform automatically determines the data type based on this logic. However, if the data type is not determined correctly, you can change it manually.
Note
Choosing the wrong input data type may reduce model accuracy or increase model size.
System uses the entire dataset to determine the data type:
- 8-bit Integer — The range of values in the dataset is from -128 to 127.
- 16-bit Integer — The range of values in the dataset is -32 768 to 32 767.
- 32-bit Floating point — At least one value in the dataset is a floating point number.
Note
If you selected Axon as the technology during the solution creation step, the only currently available option for the input data type is 32-bit floating point.