Custom Python Model Predict — Machine Learning/Python Models

Custom Python Model Predict

Machine Learning / Python Models

Runs a model you load in Python, once per sample: [y1, y2, …] = predict(model, [u1, u2, …]), where model = load_model(arguments) comes from a Python module you name. It follows the contract of MATLAB’s Custom Python Model Predict block, so the same module runs in both.

The module

• load_model(...) – called once when the run starts, with the Load Model Arguments; returns the model.
• predict(model, inputs) – called once per sample with the list of input arrays; returns a list with one array per output port.
• reset_model(model) – optional; see Notes.

Ports

• Inputs – u1, u2, … in port order, one array each in predict’s list: cast to that input’s entry of Input Data Types and brought to its entry of Input Ranks. A vector (a column or a row) of n is a 1-D array of n, as a Simulink 1-D signal is; an m×n matrix is an (m, n) array. One input by default; the count is user-editable.
• Outputs – y1, y2, … in the order predict returns them: a single value is [1,1], a 1-D array of k is [k,1], a 2-D array keeps its shape. Numbers and booleans only. One output by default; the count is user-editable and must match what predict returns.

Parameters

• Python Module Path – the .py file defining load_model and predict. Its folder is put on Python’s path, so it may import files beside it. Empty runs a built-in example whose predict returns every input times a gain (2 unless load_model is given one), so it needs as many outputs as inputs. A relative path is read from the directory the app was started in.
• Load Model Arguments – what load_model is called with, comma-separated, as in MATLAB’s block: a number, Inf, NaN, true or false, or a bracketed matrix of numbers ([2 3] is a 1-D array, [1;2] and [1 2;3 4] are 2-D, [] is empty), crosses as a number or a numpy array; anything else crosses as its own text. MATLAB evaluates wider numeric expressions (pi, 1:3, 2*3): write those as their values here. Empty calls load_model with no arguments.
• Input Data Types – one numpy type per input, comma-separated: float32 (the default, and what an empty entry means), float64, float16, float, int8, int16, int32, int64, int, uint8, uint16, uint32 or uint64. A shorter list is padded with float32. The cast rounds (float32) or truncates (the integer types).
• Input Ranks – one rank per input, comma-separated: inherit (the default, and what an empty entry means) – 1 for a vector or a scalar, 2 for a matrix – or a whole number, reached by adding or removing TRAILING dimensions of size 1 (so a vector at rank 2 is an (n, 1) column). A shorter list is padded with inherit.
• Preprocessing File – optional: a Python file defining preprocess(model, inputs), which receives the input list and returns the list predict gets.
• Postprocessing File – optional: a Python file defining postprocess(model, outputs), which receives predict’s list and returns the list the outputs take.
• Sampling Time (s) – zero or less inherits the solver's rate; a positive value runs the block at that period.

Code export

Python only – the generated module imports the same Python module and calls it exactly as the simulation does. The other nine targets (MATLAB, Java, Rust, C, C++, VHDL, Verilog, SystemVerilog, PLC Structured Text) are not supported: none can carry a Python interpreter, so an export to any of them fails loudly for this block.

Simulink bridge

Both directions, to statsPycoex/Custom Python Model Predict. Python Module Path ↔ PythonModulePath, Load Model Arguments ↔ LoadModelArgs, Preprocessing File ↔ PreprocessingFilePath, Postprocessing File ↔ PostprocessingFilePath. The port counts, Input Data Types and Input Ranks cross together as InputTable and OutputTable, one row per port. Input and output permutations and variable-size outputs have no setting here and are reported, not crossed. A block on the built-in example has no module to name and is reported on export. Sampling Time (s) → SampleTime, as on every block.

Notes

• Discrete by nature: the model is called once per sample, never integrated.
• Before the first sample, as in MATLAB’s block, predict runs ONCE on inputs of all ones – the output sizes are taken from it and must not change during the run – and then the module’s reset_model(model) runs, if it has one. A model that keeps state should reset it there; without reset_model that first call’s effect stays.
• Needs the Python runtime and numpy; a module that fails to load or to predict is reported once, with the block’s path, and stops the run.