pytorch2ltspice 0.1.3

Convert PyTorch neural network models into LTspice subcircuits (.subckt).

pytorch2ltspice

pytorch2ltspice converts PyTorch neural network models into LTspice-compatible subcircuits (.subckt). By combining it with LTspicePowerSim, users can implement AI-based controllers directly in power electronics circuits such as DC-DC converters, inverters, and motor drivers.
This repository also provides example code where a neural network controls the PWM of a BUCK regulator, trained with behavior imitation and PPO.

---

📌 Features

• Converts PyTorch nn.Sequential models to LTspice-compatible .subckt format
• Supported layers:
  • Linear:
    • nn.Linear
  • Activations:
    • nn.ReLU
    • nn.Sigmoid
    • nn.Tanh
  • Cells:
    • nn.RNNCell
    • nn.GRUCell
    • nn.LSTMCell
• Outputs a netlist using behavioral voltage sources (B elements)
• Recurrent cells are implemented with .machine blocks (LO/LATCH/HI states, CLK pin auto-added)
• Auto-generates LTspice node names (NNIN1, NNIN2, ..., NNOUT1, ...)
• Easy integration into LTspice testbenches
• Utilities for signal generation, model scaffolding, and sampling (siggen, modelgen, sampling)
• Several example models are included:
  • MLP: Linear → ReLU → Linear → ReLU → Linear → Sigmoid
  • GRUCell: GRUCell → Linear → Tanh
  • LSTMCell: LSTMCell → Linear → Tanh
  • RNNCell: RNNCell → Linear → Tanh
  • Hybrid: Linear → ReLU → (GRUCell/LSTMCell) → Linear → Tanh
  • Multi-cell LSTM: stacked LSTMCell layers

---

🧠 Motivation

Neural networks trained in Python (with PyTorch) can now be exported and tested directly in LTspice circuit simulations.
This allows for:

• Closed-loop simulation with NN controllers
• Verification of inference logic inside switching power supplies
• Observation of behavior under nonlinear and dynamic conditions

---

🚀 Installation

Option A: Install from PyPI (recommended)

pip install pytorch2ltspice

Option B: Install from source (GitHub)

git clone https://github.com/kosokno/pytorch2ltspice.git
cd pytorch2ltspice
pip install -e .

---

⚡ Quick Start

1. Define a model in PyTorch

import torch.nn as nn

model = nn.Sequential(
    nn.Linear(20, 32),
    nn.ReLU(),
    nn.Linear(32, 16),
    nn.ReLU(),
    nn.Linear(16, 1)
)
model.eval()

2. Export as LTspice .subckt file

from pytorch2ltspice import export_model_to_ltspice

export_model_to_ltspice(
    model,
    filename="TEST_MODEL_SUBCKT.SP",
    subckt_name="TESTACTORSUBCKT"
)

3. Include it in LTspice

• Add the following directive in LTspice:
  • .include TEST_MODEL_SUBCKT.SP
• Wire NNIN* pins to your signals and read NNOUT* as the inference output.

---

🧰 Utilities

Helper utilities are available under pytorch2ltspice.utils:

• siggen: signal generators for LTspice testbenches
• modelgen: minimal model scaffolding for experiments
• sampling: simple data sampling helpers

Note: example code now uses these utilities. Developer tools moved from tools/modelgen to tools/utils_test.

---

📂 Output Example

The resulting LTspice subcircuit will look like:

.SUBCKT TESTACTORSUBCKT NNIN1 NNIN2 ... NNIN20 NNOUT1
* LAYER 1: LINEAR
B1_1 L1_1 0 V=V(NNIN1)*(-0.179081)+V(NNIN2)*(-0.068428)+...
...
* ACTIVATION LAYER 1: RELU
B_ACT1_1 L_ACT1_1 0 V=(IF(V(L1_1)>0,V(L1_1),0))
...
B_OUT NNOUT1 0 V=V(L_ACT2_1)
.ENDS TESTACTORSUBCKT

---

✨ Training Example

NN Controlled Voltage Mode Buck

---

📄 License

MIT License

---

🧩 Related Projects

• 🔗 LTspicePowerSim:
  A Simulink-like power electronics simulation environment built on LTspice,