Label Encoder — Machine Learning/Preprocessing

Label Encoder

Machine Learning / Preprocessing

Turns a text label into the class code a model expects: y = the position of u in the label list, counting from Index Base. Text that is in the list nowhere comes out as Unknown Code. This is the block that lets a name – a mode, a state, a category read off a bus – drive arithmetic.

Ports

• Input – the label u (str, ICoreString), one value. Only text: a number arriving here has no label to be, and the one way to change a signal's type is a Data Type Conversion in the wire.
• Output – the code y (i32, ICoreInt32), one value. A code is a whole number and it is frequently negative – -1 is the usual Unknown Code – so it is a signed integer.

Parameters

• Labels – the fitted label list, separated by semicolons: Idle;Run;Fault. Order is what defines the codes, so it must be the order the model was fitted with. Surrounding spaces are kept, because a label may legitimately have one; a label may not itself contain a semicolon. An empty list is allowed and means every input is unknown.
• Unknown Code – the code for text that matches no label. Whole number, and it is not shifted by Index Base: it names something that is not a class.
• Index Base – what the first label is called:
  • Zero-based (PyTorch, numpy) – the first label is code 0.
  • One-based (MATLAB) – the first label is code 1.
• Sampling Time (s) – zero or less inherits the solver's rate; a positive value runs the block at that period.

Matching is exact, byte for byte

A label matches only text that is identical to it – same letters, same case, same spaces. There is deliberately no case-insensitive setting: the six languages this block exports to disagree about what lowercasing means the moment the text stops being plain English (one of them makes the text longer), and a rule that held in five targets and not the sixth would be worse than no rule. To accept two spellings, list them both; both codes can then be treated as one downstream.

If the same label is listed twice, the first position wins – the lowest code. That is the tie rule Argmax Decision and One-Hot Encoder already use, so a chain of the three never has to be read twice.

Code export

C, C++, Python, MATLAB, Java and Rust all carry this block, and all six answer the same code as the simulation for every input. The label list is baked into the generated code at export time rather than being offered as a tunable parameter: it is what the comparison IS, not a number to retune afterwards.

VHDL, Verilog, SystemVerilog and PLC Structured Text do not carry a text signal, so an export to one of them stops and names this block and the reason rather than emitting something that does not run.

Simulink bridge

None – Support::None. Simulink's String library compares, searches and converts text, and none of it maps a fitted label set onto a class code; its Statistics blocks take a fitted model object rather than a parameter set, which no bridge parameter can carry. The block is reported as unsupported rather than silently dropped.

Notes

• Algebraic, with no state: the output depends only on the current input.
• An unconnected input is empty text. That is unknown unless the empty string is itself one of the labels, which is allowed – Idle;;Fault makes the empty string code 1.
• Text is carried and compared to 256 bytes, the limit every text signal in this library is cut at, so a label longer than that can never be matched.