_        __                                                     _            
 (_)      / _|                                                   (_)           
  _ _ __ | |_ ___ _ __ ___ _ __   ___ ___         ___ _ __   __ _ _ _ __   ___ 
 | | '_ \|  _/ _ \ '__/ _ \ '_ \ / __/ _ \  __   / _ \ '_ \ / _` | | '_ \ / _ \
 | | | | | ||  __/ | |  __/ | | | (_|  __/ |__| |  __/ | | | (_| | | | | |  __/
 |_|_| |_|_| \___|_|  \___|_| |_|\___\___|       \___|_| |_|\__, |_|_| |_|\___|
                                                             __/ |             
                                                            |___/              

Inference-Engine

Table of contents

• Overview
• Downloading, Building and testing
• Examples
• Documentation

Overview

Inference-Engine supports research in concurrent, large-batch inference and training of deep, feed-forward neural networks. Inference-Engine targets high-performance computing (HPC) applications with performance-critical inference and training needs. The initial target application is in situ training of a cloud microphysics model proxy for the Intermediate Complexity Atmospheric Research (ICAR) model. Such a proxy must support concurrent inference at every grid point at every time step of an ICAR run. For validation purposes, Inference-Engine also supports the export and import of neural networks to and from Python by the companion package nexport.

The features of Inference-Engine that make it suitable for use in HPC applications include

1. Implementation in Fortran 2018.
2. Exposing concurrency via

• Elemental, implicitly pure inference procedures,
• An elemental and implicitly pure activation strategy, and
• A pure training subroutine,

2. Gathering network weights and biases into contiguous arrays for efficient memory access patterns, and
3. User-controlled mini-batch size facilitating in situ training at application runtime.

Making Inference-Engine's infer functions and train subroutines pure facilitates invoking those procedures inside Fortran do concurrent constructs, which some compilers can offload automatically to graphics processing units (GPUs). The use of contiguous arrays facilitates spatial locality in memory access patterns. User control of mini-batch size facilitates in-situ training at application runtime.

The available optimizers for training neural networks are

1. Stochastic gradient descent
2. Adam (recommended)

Build and Test

With the Fortran Package Manager (fpm) and a recent version of a Fortran compiler installed, enter one of the commmands below to build the Inference-Engine library and run the test suite:

GNU (gfortran)

fpm test

Intel (ifx)

fpm test --compiler ifx

Experimental: Automatic offloading of do concurrent to GPUs

This capability is under development with the goal to facilitate GPU automatic offloading via the following command:

fpm test --compiler ifx --flag "-fopenmp-target-do-concurrent -qopenmp -fopenmp-targets=spir64"

LLVM (flang-new)

Support for LLVM flang-new is under development and currently requires building flang-new from source with assumed-rank support enabled:

fpm test --compiler flang-new --flag "-mmlir -allow-assumed-rank"

A script that might help with building flang-new from source is in the handy-dandy repository.

NAG (nagfor) -- under development

fpm test --compiler nagfor --flag -fpp

HPE (crayftn.sh) -- under development

Support for the Cray Compiler Environment (CCE) Fortran compiler is under development. Building with the CCE ftn compiler wrapper requires an additional trivial wrapper. With a shell script named crayftn.sh of the following form in your PATH

#!/bin/bash

ftn "$@"

execute the following command:

fpm test --compiler crayftn.sh

Examples

The example subdirectory contains demonstrations of several intended use cases.

Configuring a Training Run

To see the format for a JSON configuration file that defines the hyperparameters and a new network configuration for a training run, execute the provided training-configuration output example program:

% ./build/run-fpm.sh run --example print-training-configuration
Project is up to date
 {
     "hyperparameters": {
         "mini-batches" : 10,
         "learning rate" : 1.50000000,
         "optimizer" : "adam"
     }
 ,
     "network configuration": {
         "skip connections" : false,
         "nodes per layer" : [2,72,2],
         "activation function" : "sigmoid"
     }
 }

As of this writing, the JSON file format is fragile. Because an Intel ifx compiler bug prevents using our preferred JSON interface, rojff, Inference-Engine currently uses a very restricted JSON subset written and read by the sourcery utility's string_t type-bound procedures. For this to work, it is important to keep input files as close as possible to the exact form shown above. In particular, do not split, combine or reorder lines. Adding or removing whitespace should be ok.

Documentation

Please see the Inference-Engine GitHub Pages site for HTML documentation generated by ford.