Neural network layer/op support¶

Below is a list of all TFLite 'ops' (operations or neural network layer types) that are supported by the inference engine. The main data-type of the inference engine is quantized 8-bit integers ('INT8'), which should be used for best performance. Additionally, many ops, including the most common ones, are also available in quantized 16-bit ('INT16'). Some supporting ops are also available in non-quantized 32-bit integer ('INT32') or 32-bit floating-point ('FLOAT32') mode. These supporting ops are typically only used for simple index or shape computations.

For some ops below there are certain conditions for which a certain data-type is supported. This can depend on the op parameters for example. Those conditions are mentioned in the 'Notes' column and/or will be checked at run-time. More details for each TFLite op type can be found on the TFLite MLIR website.

We group the ops in these types to make the long table a bit easier to parse:

• NN layer: A neural network layer, typically only available in quantized INT8
• NN activation: A neural network activation function
• Math op: A mathematical support op, typically only available in FLOAT32
• Support op: Any other support operation, used e.g. for indexing or reshaping