Skip to content

Transformer Regime Plugin (TorchScript)

This section describes a TorchScript-backed regime detector plugin. It requires libtorch.

Overview

  • Load a TorchScript model (.pt) at plugin initialization.
  • Compute features from bars/ticks in C++.
  • Run the model to produce regime probabilities.
  • Map probabilities to RegimeFlow regimes.

Expected Config

regime:
  detector: transformer_torchscript
  params:
    model_path: /path/to/regime_transformer.pt
    feature_window: 120
    feature_dim: 8

Build Requirements

  • libtorch (C++ API for PyTorch)
  • CMake find_package(Torch) integration

Implementation

  • Implementation lives in:
  • examples/plugins/transformer_regime/transformer_torchscript_detector.cpp

  • examples/plugins/transformer_regime/transformer_torchscript_detector.h

  • Normalize features identically to the Python training pipeline.

Notes

  • TorchScript model should output logits for 4 regimes (bull/neutral/bear/crisis).
  • Ensure deterministic inference (set eval mode, disable grads).

The plugin loads a TorchScript module, runs inference on a [1, window, feature_dim] tensor, and maps the probabilities to RegimeFlow regime labels.

This file documents the integration; implementation can be added once libtorch is available.