TorchFort: An Online Deep Learning Interface for HPC programs on NVIDIA GPUs

These pages contain the documentation for TorchFort, an online deep learning interface for HPC programs.

TorchFort is a DL training and inference interface for HPC programs implemented using LibTorch, the C++ backend used by the PyTorch framework. The goal of this library is to help practitioners and domain scientists to seamlessly combine their simulation codes with Deep Learning functionalities available within PyTorch. This library can be invoked directly from Fortran or C/C++ programs, enabling transparent sharing of data arrays to and from the DL framework all contained within the simulation process (i.e., no external glue/data-sharing code required). The library can directly load PyTorch model definitions exported to TorchScript and implements a configurable training process that users can control via a simple YAML configuration file format. The configuration files enable users to specify optimizer and loss selection, learning rate schedules, and much more.

Please contact us or open a GitHub issue if you are interested in using this library in your own solvers and have questions on usage and/or feature requests.

Table of Contents

• Installation
  • Docker Installation
  • CMake Installation
  • Build Documentation
  • Directory Structure
• Usage Guide
  • Supervised Learning
    • Creating a model
    • Run a Training Step
    • Run Inference to generate predicted output
    • Checkpoint/Restart
  • Reinforcement Learning
    • Creating an RL system
    • Replay Buffer Management
    • Generating Action Predictions
    • Training Step
    • Checkpoint/Restart
• TorchFort API
  • TorchFort Configuration Files
    • Common
      • General Properties
      • Optimizer Properties
      • Learning Rate Schedule Properties
    • Supervised Learning
      • Model Properties
      • Loss Properties
    • Reinforcement Learning
      • Reinforcement Learning Training Algorithm Properties
      • Replay Buffer Properties
      • Action Properties
      • Policy and Critic Properties
      • Learning Rate Schedule Properties
  • TorchFort C API
    • General
      • Types
      • Global Context Settings
    • Supervised Learning
      • Model Creation
      • Model Training/Inference
      • Model Management
      • Weights and Biases Logging
    • Reinforcement Learning
      • Off-Policy Algorithms
      • On-Policy Algorithms
  • TorchFort Fortran API
    • General
      • Types
      • Global Context Settings
    • Supervised Learning
      • Model Creation
      • Model Training/Inference
      • Model Management
      • Weights and Biases Logging
    • Reinforcement Learning
      • Off-Policy Algorithms
      • On-Policy Algorithms
• Extras
  • Weights and Biases Support
    • Add Custom Metrics Reporting to Application
    • Set up Environment
    • Start Background Watcher Process
    • Start Your TorchFort Application

Indices and tables

• Index

• Module Index

• Search Page