跳到主要内容

onnx_exporter

Module: GBC.gyms.isaaclab_45.lab_tasks.utils.wrappers.rsl_rl.onnx_exporter

🎯 Overview​

The ONNX exporter provides seamless conversion of BERT-form transformer PPO architectures to ONNX format for efficient deployment. This module leverages the latest ONNX library to export trained policies, enabling cross-platform inference and optimized production deployment.

🚀 Key Benefits​

  • 🌐 Cross-Platform Deployment: Run trained policies on various hardware and software platforms
  • ⚡ Optimized Inference: ONNX runtime provides highly optimized execution
  • 🔧 Framework Agnostic: Export from PyTorch to platform-independent format
  • 📱 Edge Deployment: Enable deployment on edge devices and embedded systems

📦 Public Functions​

export_policy_as_onnx()​

def export_policy_as_onnx(
    actor_critic: object, 
    path: str, 
    normalizer: object | None = None, 
    filename: str = "policy.onnx", 
    verbose: bool = False
) -> None

Purpose: Export a trained actor-critic policy to ONNX format for deployment.

Parameters​

  • actor_critic (object): The trained actor-critic torch module to export
  • path (str): Directory path where the ONNX file will be saved
  • normalizer (object | None): Empirical normalizer module. If None, Identity normalization is used
  • filename (str): Name of the exported ONNX file (default: "policy.onnx")
  • verbose (bool): Whether to print detailed export information (default: False)

Functionality​

  • 📁 Directory Creation: Automatically creates output directory if it doesn't exist
  • 🧠 Model Processing: Handles both recurrent (LSTM) and non-recurrent transformer architectures
  • 🔄 Reference Support: Automatically detects and exports reference observation capabilities
  • ⚙️ Optimization: Uses appropriate ONNX opset versions for best compatibility

Usage Example​

from GBC.gyms.isaaclab_45.lab_tasks.utils.wrappers.rsl_rl import export_policy_as_onnx

# Export trained policy
export_policy_as_onnx(
    actor_critic=trained_policy,
    path="./exported_models",
    normalizer=obs_normalizer,
    filename="robot_policy.onnx",
    verbose=True
)

🏗️ Internal Implementation​

_OnnxPolicyExporter Class​

Purpose: Private helper class that handles the actual ONNX conversion process.

Key Features​

🧠 Architecture Detection

  • Automatically detects recurrent vs. non-recurrent models
  • Handles BERT-form transformer architectures
  • Supports reference observation embeddings

📊 Input/Output Handling

  • Standard Mode: obs → actions
  • Reference Mode: obs, ref_obs, ref_mask → actions
  • Recurrent Mode: obs, h_in, c_in → actions, h_out, c_out

⚙️ ONNX Configuration

  • Opset Version: Uses opset 14 for non-recurrent, opset 11 for recurrent models
  • Dynamic Axes: Configurable for flexible batch sizes
  • Export Parameters: Includes all learned weights and biases

Supported Model Types​

🔄 Non-Recurrent Models (Transformer-based)

# Input specification
input_names = ["obs"]  # Standard observations
if reference_mode:
    input_names += ["ref_obs", "ref_mask"]  # Reference observations + mask

# Output specification  
output_names = ["actions"]

🔁 Recurrent Models (LSTM-based)

# Input specification
input_names = ["obs", "h_in", "c_in"]  # Observations + hidden states

# Output specification
output_names = ["actions", "h_out", "c_out"]  # Actions + updated states

🎯 Export Process Flow​

graph TD
    A[Trained Policy] --> B[_OnnxPolicyExporter]
    B --> C{Architecture Type?}
    C -->|Transformer| D[Configure Non-Recurrent Export]
    C -->|LSTM| E[Configure Recurrent Export]
    D --> F{Reference Observations?}
    F -->|Yes| G[Multi-Input ONNX Graph]
    F -->|No| H[Single-Input ONNX Graph]
    E --> I[LSTM ONNX Graph]
    G --> J[Export ONNX File]
    H --> J
    I --> J
    J --> K[Deployment-Ready Model]

🔧 Technical Specifications​

ONNX Compatibility​

  • Framework: PyTorch → ONNX conversion
  • Opset Versions: 11 (recurrent) / 14 (non-recurrent)
  • Data Types: FP32 tensors
  • Batch Dimension: Configurable dynamic axes

Model Architecture Support​

  • ✅ BERT-form Transformers: Full support for attention-based policies
  • ✅ Reference Observations: Multi-modal input handling
  • ✅ LSTM Networks: Recurrent policy support
  • ✅ Observation Normalization: Embedded preprocessing

Deployment Targets​

  • 🖥️ ONNX Runtime: CPU/GPU inference
  • 📱 Mobile Platforms: iOS/Android deployment
  • 🔧 Edge Devices: Embedded systems and microcontrollers
  • ☁️ Cloud Services: Scalable inference endpoints

💡 Best Practices​

🎯 Export Optimization​

  1. Model Preparation: Ensure model is in evaluation mode before export
  2. Device Placement: Export process automatically moves model to CPU
  3. Input Shapes: Zero tensors are used to trace the computation graph
  4. Normalizer Integration: Include observation normalizers in the export

📊 Quality Assurance​

  1. Validation: Test exported ONNX model against original PyTorch version
  2. Performance: Benchmark inference speed on target deployment platform
  3. Compatibility: Verify ONNX runtime version compatibility
  4. Precision: Check numerical precision matches original model

🚀 Deployment Workflow​

# 1. Export trained policy
export_policy_as_onnx(policy, "./models", normalizer, "robot.onnx")

# 2. Load in ONNX runtime
import onnxruntime as ort
session = ort.InferenceSession("./models/robot.onnx")

# 3. Run inference
outputs = session.run(None, {"obs": observation_data})
actions = outputs[0]

🔗 Related Components​

  • Actor-Critic Models: Source models for export
  • Observation Normalizers: Preprocessing components
  • ONNX Runtime: Deployment inference engine
  • Transformer Architectures: BERT-form policy networks