Skip to content
/ mlia Public

The ML Inference Advisor (MLIA) helps AI developers design and optimize neural network models for efficient inference on Arm® targets. MLIA provides insights on how the ML model will perform on Arm early in the model development cycle and advice operator compatibility, performance analysis and model optimization.

0 stars 0 forks Branches Tags Activity
Star
Notifications You must be signed in to change notification settings

arm/mlia

BranchesTags

Repository files navigation

ML Inference Advisor - Introduction

The ML Inference Advisor (MLIA) helps AI developers design and optimize neural network models for efficient inference on Arm® targets (see supported targets). MLIA provides insights on how the ML model will perform on Arm early in the model development cycle. By passing a model file and specifying an Arm hardware target, users get an overview of possible areas of improvement and actionable advice. The advice can cover operator compatibility, performance analysis and model optimization (e.g. pruning and clustering). With the ML Inference Advisor, we aim to make the Arm ML IP accessible to developers at all levels of abstraction, with differing knowledge on hardware optimization and machine learning.

Inclusive language commitment

This product conforms to Arm's inclusive language policy and, to the best of our knowledge, does not contain any non-inclusive language.

If you find something that concerns you, email [email protected].

Releases

Release notes can be found in MLIA releases.

Getting support

In case you need support or want to report an issue, give us feedback or simply ask a question about MLIA, please send an email to [email protected].

Alternatively, use the AI and ML forum to get support by marking your post with the MLIA tag, or tag the @mlia team directly for assistance.

Reporting vulnerabilities

Information on reporting security issues can be found in Reporting vulnerabilities.

License

ML Inference Advisor is licensed under Apache License 2.0 unless otherwise indicated. This project contains software under a range of permissive licenses, see LICENSES.

Trademarks and copyrights

  • Arm®, Arm® Ethos™-U, Arm® Cortex®-A, Arm® Cortex®-M, Arm® Corstone™ are registered trademarks or trademarks of Arm® Limited (or its subsidiaries) in the U.S. and/or elsewhere.
  • TensorFlow™ is a trademark of Google® LLC.
  • Keras™ is a trademark by François Chollet.
  • Linux® is the registered trademark of Linus Torvalds in the U.S. and elsewhere.
  • Python® is a registered trademark of the PSF.
  • Ubuntu® is a registered trademark of Canonical.
  • Microsoft and Windows are trademarks of the Microsoft group of companies.

General usage

Prerequisites and dependencies

It is recommended to use a virtual environment for MLIA installation, and a typical setup requires:

  • Ubuntu® 22.04.5 LTS (other OSs may work, the ML Inference Advisor has been tested on this one specifically)
  • Python® >= 3.9
  • libpython3.9-dev (part of python3.9-dev)

Installation

MLIA can be installed with pip using the following command:

pip install mlia

It is highly recommended to create a new virtual environment for the installation.

First steps

After the installation, you can check that MLIA is installed correctly by opening your terminal, activating the virtual environment and typing the following command that should print the help text:

mlia --help

The ML Inference Advisor works with sub-commands, i.e. in general a command would look like this:

mlia [sub-command] [arguments]

Where the following sub-commands are available:

  • "check": perform compatibility or performance checks on the model
  • "optimize": apply specified optimizations

Detailed help about the different sub-commands can be shown like this:

mlia [sub-command] --help

The following sections go into further detail regarding the usage of MLIA.

Sub-commands

This section gives an overview of the available sub-commands for MLIA.

check

compatibility

Lists the model's operators with information about their compatibility with the specified target.

Examples:

# List operator compatibility with Ethos-U55 with 256 MAC
mlia check ~/models/mobilenet_v1_1.0_224_quant.tflite --target-profile ethos-u55-256

# List operator compatibility with Cortex-A
mlia check ~/models/mobilenet_v1_1.0_224_quant.tflite --target-profile cortex-a

# Get help and further information
mlia check --help

performance

Estimates the model's performance on the specified target and prints out statistics.

Examples:

# Use default parameters
mlia check ~/models/mobilenet_v1_1.0_224_quant.tflite \
    --target-profile ethos-u55-256 \
    --performance

# Explicitly specify the target profile and backend(s) to use
# with --backend option
mlia check ~/models/ds_cnn_large_fully_quantized_int8.tflite \
    --target-profile ethos-u65-512 \
    --performance \
    --backend "vela" \
    --backend "corstone-300"

# Get help and further information
mlia check --help

optimize

This sub-command applies optimizations to a Keras model (.h5 or SavedModel) or a TensorFlow Lite model and shows the performance improvements compared to the original unoptimized model.

There are currently three optimization techniques available to apply:

  • pruning: Sets insignificant model weights to zero until the specified sparsity is reached.
  • clustering: Groups the weights into the specified number of clusters and then replaces the weight values with the cluster centroids.

More information about these techniques can be found online in the TensorFlow documentation, e.g. in the TensorFlow model optimization guides.

Examples:

# Custom optimization parameters: pruning=0.6, clustering=16
mlia optimize ~/models/ds_cnn_l.h5 \
    --target-profile ethos-u55-256 \
    --pruning \
    --pruning-target 0.6 \
    --clustering \
    --clustering-target 16

# Get help and further information
mlia optimize --help

Note: A Keras model (.h5 or SavedModel) is required as input to perform pruning and clustering.

rewrite

Replaces certain subgraph/layer of the pre-trained model with candidates from the rewrite library, with or without training using a small portion of the training data, to achieve local performance gains.

The following rewrites are supported:

  • fully-connected - replaces a subgraph with a fully connected layer
  • fully-connected-sparsity - replaces a subgraph with a pruned M:N sparse fully connected layer
  • fully-connected-unstructured-sparsity - replaces a subgraph with an unstructured pruned fully connected layer
  • fully-connected-clustering - replaces a subgraph with a clustered fully connected layer
  • conv2d - replaces a subgraph with a conv2d layer
  • conv2d-sparsity - replaces a subgraph with a pruned M:N sparse conv2d layer
  • conv2d-unstructured-sparsity - replaces a subgraph with an unstructured pruned conv2d layer
  • conv2d-clustering - replaces a subgraph with a clustered conv2d layer
  • depthwise-separable-conv2d - replaces a subgraph with a depthwise separable conv2d layer
  • depthwise-separable-conv2d-sparsity - replaces a subgraph with a pruned M:N sparse depthwise separable conv2d layer
  • depthwise-separable-conv2d-unstructured-sparsity - replaces a subgraph with an unstructured pruned depthwise separable conv2d layer
  • depthwise-separable-conv2d-clustering - replaces a subgraph with a clustered depthwise separable conv2d layer

Note: A TensorFlow Lite model is required as input to perform a rewrite.

Examples:

# Rewrite Example 1
mlia optimize ~/models/ds_cnn_large_fp32.tflite \
    --target-profile ethos-u55-256 \
    --rewrite \
    --dataset input.tfrec \
    --rewrite-target fully-connected \
    --rewrite-start MobileNet/avg_pool/AvgPool \
    --rewrite-end MobileNet/fc1/BiasAdd

# Rewrite Example 2
mlia optimize ~/models/ds_cnn_large_fp32.tflite \
    --target-profile ethos-u55-256 \
    --rewrite \
    --dataset input.tfrec \
    --rewrite-target conv2d-clustering \
    --rewrite-start model/re_lu_9/Relu \
    --rewrite-end model/re_lu_10/Relu

Random Dataset

The dataset flag is optional. If you do not provide a dataset, then the rewrite will occur using random data to give the user an idea of the performance benefits of the rewrite on the model.

Conv2d Rewrites

For conv2d rewrites, the conv2d layer parameters are calculated as followed:

We first assume that valid (no) padding will be used, we calculate the conv2d parameters using the following formulae:

Kernel size: set by the user, defaults to 3x3

Output filters = $output_shape[-1]$

$stride[0] = input_shape[0] / output_shape[0]$ (rounded to nearest integer)

$stride[1] = input_shape[1] / output_shape[1]$ (rounded to nearest integer)

The input and output shapes are then calculated using the following formulae:

$output_shape[0] = \lfloor(input_shape[0] - kernel_size[0]) / stride[0] \rfloor + 1$

$output_shape[1] = \lfloor(input_shape[1] - kernel_size[1]) / stride[1] \rfloor + 1$

If these resulting sizes do not match the desired output shape, we set the padding to 'same' such that they match it.

This introduces some constraints into the size of the kernel that can be used with the rewrite subgraph to produce the desired output shape. The user should be aware of these formulae when performing rewrites.

Optimization Profiles

Training parameters for rewrites can be specified.

There are a number of predefined profiles for rewrites. Some examples of these are shown below:

Name Batch Size LR Show Progress Steps LR Schedule Num Procs Num Threads Checkpoints
optimization 32 1e-3 True 48000 "cosine" 1 0 None
Name Batch Size LR Show Progress Steps LR Schedule Num Procs Num Threads Checkpoints Num Clusters Cluster Centroids Init
optimization-fully-connected-clustering 32 1e-3 True 48000 "cosine" 1 0 None 16 "CentroidInitialization.LINEAR"
Name Batch Size LR Show Progress Steps LR Schedule Num Procs Num Threads Checkpoints Sparsity M Sparsity N
optimization-fully-connected-pruning 32 1e-3 True 48000 "cosine" 1 0 None 2 4
Name Batch Size LR Show Progress Steps LR Schedule Num Procs Num Threads Checkpoints Initial Sparsity End Sparsity End Step
optimization-fully-connected-unstructured-pruning 32 1e-3 True 48000 "cosine" 1 0 None 0.25 0.5 48000
Name Batch Size LR Show Progress Steps LR Schedule Num Procs Num Threads Checkpoints Num Clusters Cluster Centroids Init Activation Kernel Size
optimization-conv2d-clustering 32 1e-3 True 48000 "cosine" 1 0 None 16 "CentroidInitialization.LINEAR" "relu" 3x3
Name Batch Size LR Show Progress Steps LR Schedule Num Procs Num Threads Checkpoints Sparsity M Sparsity N Activation Kernel Size
optimization-conv2d-pruning 32 1e-3 True 48000 "cosine" 1 0 None 2 4 "relu" 3x3
Name Batch Size LR Show Progress Steps LR Schedule Num Procs Num Threads Checkpoints Initial Sparsity End Sparsity End Step Activation Kernel Size
optimization-conv2d-unstructured-pruning 32 1e-3 True 48000 "cosine" 1 0 None 0.25 0.5 48000 "relu" 3x3

The complete list of built in optimization profiles is shown below. Each profile provides training parameters and parameters specific to the rewrite.

  • optimization
  • optimization-fully-connected-clustering
  • optimization-fully-connected-pruning
  • optimization-fully-connected-unstructured-pruning
  • optimization-conv2d
  • optimization-conv2d-clustering
  • optimization-conv2d-pruning
  • optimization-conv2d-unstructured-pruning
  • optimization-depthwise-separable-conv2d
  • optimization-depthwise-separable-conv2d-clustering
  • optimization-depthwise-separable-conv2d-pruning
  • optimization-conv2d-depthwise-separable-unstructured-pruning

Note: For convolutional rewrites (e.g. optimization-conv2d-pruning). The activation function for the rewrite can be selected in the optimization profile from the following list:

  • "relu" - Standard ReLU activation function
  • "relu6" - ReLU6 activation function i.e. ReLU activation function capped at 6
  • "none" - No activation function

The user can also specify custom augmentations as part of the training parameters. An example of this can be found in the following optimization profile:

Name Batch Size LR Show Progress Steps LR Schedule Num Procs Num Threads Checkpoints Augmentations - gaussian_strength Augmentations - mixup_strength
optimization-custom-augmentation 32 1e-3 True 48000 "cosine" 1 0 None 0.1 0.1

The augmentations consist of 2 parameters: mixup strength and gaussian strength.

Augmentations can be selected from a number of pre-defined profiles (see the table below) or each individual parameter can be chosen (see optimization_custom_augmentation above for an example):

Name MixUp Strength Gaussian Strength
"none" None None
"gaussian" None 1.0
"mixup" 1.0 None
"mixout" 1.6 None
"mix_gaussian_large" 2.0 1.0
"mix_gaussian_small" 1.6 0.3

An example of using an optimization profile can be seen below:

##### An example for using optimization Profiles
mlia optimize ~/models/ds_cnn_large_fp32.tflite \
    --target-profile ethos-u55-256 \
    --optimization-profile optimization \
    --rewrite \
    --dataset input.tfrec \
    --rewrite-target fully-connected \
    --rewrite-start MobileNet/avg_pool/AvgPool \
    --rewrite-end MobileNet/fc1/BiasAdd

Custom optimization Profiles

For the custom optimization profiles, the configuration file for a custom optimization profile is passed as path and needs to conform to the TOML file format. Each optimization in MLIA has a pre-defined set of parameters which can be present in the config file. When using the built-in optimization profiles, the appropriate toml file is copied to mlia-output and can be used to understand what parameters apply for each optimization.

Example:

# for custom profiles
mlia optimize --optimization-profile ~/my_custom_optimization_profile.toml

When providing rewrite-specific parameters e.g. for clustering, the rewrite name should be specified in the toml:

For example, the following provides rewrite-specific parameters for the conv2d-clustering rewrite

[rewrite.conv2d-clustering]
num_clusters = 16
cluster_centroids_init = "CentroidInitialization.LINEAR"

Target profiles

The targets currently supported are described in the sections below. All sub-commands require a target profile as input parameter. That target profile can be either a name of a built-in target profile or a custom file. MLIA saves the target profile that was used for a run in the output directory.

The support of the above sub-commands for different targets is provided via backends that need to be installed separately, see Backend installation section.

Ethos-U

There are a number of predefined profiles for Ethos-U with the following attributes:

Profile name MAC System config Memory mode
ethos-u55-256 256 Ethos_U55_High_End_Embedded Shared_Sram
ethos-u55-128 128 Ethos_U55_High_End_Embedded Shared_Sram
ethos-u65-512 512 Ethos_U65_High_End Dedicated_Sram
ethos-u65-256 256 Ethos_U65_High_End Dedicated_Sram
ethos-u85-2048 2048 Ethos_U85_SYS_DRAM_High_2048 Dedicated_Sram
ethos-u85-1024 1024 Ethos_U85_SYS_DRAM_Mid_1024 Dedicated_Sram
ethos-u85-512 512 Ethos_U85_SYS_DRAM_Mid_512 Dedicated_Sram
ethos-u85-256 256 Ethos_U85_SYS_DRAM_Low Dedicated_Sram
ethos-u85-128 128 Ethos_U85_SYS_DRAM_Low Dedicated_Sram

Example:

mlia check ~/model.tflite --target-profile ethos-u65-512 --performance

Ethos-U is supported by these backends:

As described in section Custom target profiles, you can customize the target using the following parameters in the .toml files:

  • mac: number of MACs [256, 512]
  • memory_mode: [SRAM Only, Shared SRAM, Dedicated SRAM]
  • system_config: name of the system configuration. For Vela backend, it's defined in vela.ini.
  • config: for the Vela backend - the path to Vela configuration file, passed in the --config argument. If not given, uses the builtin path: mlia/resources/vela/vela.ini

Cortex-A

DEPRECATION WARNING The cortex-a target profile uses the deprecated Arm NN TensorFlow Lite Delegate backend which will be removed in the next major release.

The profile cortex-a can be used to get the information about supported operators for Cortex-A CPUs when using the Arm NN TensorFlow Lite Delegate. Please, find more details in the section for the corresponding backend.

TOSA

DEPRECATION WARNING The tosa target profile uses the deprecated TOSA Checker backend.

The target profile tosa can be used for TOSA compatibility checks of your model. It requires the TOSA Checker backend. Please note that TOSA is currently only available for x86 architecture.

For more information, see TOSA Checker's:

Custom target profiles

For the custom target profiles, the configuration file for a custom target profile is passed as path and needs to conform to the TOML file format. Each target in MLIA has a pre-defined set of parameters which need to be present in the config file. When using the built-in target profiles, the appropriate toml file is copied to mlia-output and can be used to understand what parameters apply for each target.

Example:

# for custom profiles
mlia ops --target-profile ~/my_custom_profile.toml sample_model.tflite

Backend installation

The ML Inference Advisor is designed to use backends to provide different metrics for different target hardware. Some backends come pre-installed, but others can be added and managed using the command mlia-backend, that provides the following functionality:

  • install
  • uninstall
  • list

Examples:

# List backends installed and available for installation
mlia-backend list

# Install Corstone-300 backend for Ethos-U
mlia-backend install Corstone-300 --path ~/FVP_Corstone_SSE-300/

# Uninstall the Corstone-300 backend
mlia-backend uninstall Corstone-300

# Get help and further information
mlia-backend --help

Note: Some, but not all, backends can be automatically downloaded, if no path is provided.

Available backends

This section lists available backends. As not all backends work on any platform the following table shows some compatibility information:

Backend Linux Windows Python
Arm NN TensorFlow Lite Delegate x86_64 and AArch64 Windows 10 Python>=3.8
Corstone-300 x86_64 and AArch64 Not compatible Python>=3.8
Corstone-310 x86_64 and AArch64 Not compatible Python>=3.8
Corstone-320 x86_64 and AArch64 Not compatible Python>=3.8
TOSA checker x86_64 (manylinux2014) Not compatible 3.7<=Python<=3.9
Vela x86_64 and AArch64 Windows 10 Python~=3.7

Arm NN TensorFlow Lite Delegate

DEPRECATION WARNING This backend is deprecated and will be removed in the next major release. The Arm NN TensorFlow Lite Delegate backend relies on an unmaintained project and is no longer actively supported.

This backend provides general information about the compatibility of operators with the Arm NN TensorFlow Lite Delegate for Cortex-A. It comes pre-installed.

For version 23.05 the classic delegate is used.

For more information see:

Corstone-300

Corstone-300 is a backend that provides performance metrics for systems based on Arm® Cortex™-M55 processor and Arm® Ethos™-U NPU (Arm® Ethos™-U55 NPU or Arm® Ethos™-U65 NPU). It is only available on the Linux platform. Examples:

# Download and install Corstone-300 automatically
mlia-backend install Corstone-300
# Point to a local version of Corstone-300 installed using its installation script
mlia-backend install Corstone-300 --path YOUR_LOCAL_PATH_TO_CORSTONE_300

For further information about Corstone-300 please refer to: https://developer.arm.com/Processors/Corstone-300

Corstone-310

Corstone-310 is a backend that provides performance metrics for systems based on Arm® Cortex™-M85 processor and Ethos-U (Arm® Ethos™-U55 NPU or Arm® Ethos™-U65 NPU).

Corstone-320

Corstone-320 is a backend that provides performance metrics for systems based on Arm® Cortex™-M85 processor and Arm® Ethos™-U85 NPU.

TOSA Checker

DEPRECATION WARNING This backend is deprecated. The TOSA Checker backend relies on an unmaintained project and is no longer actively supported.

The TOSA Checker backend provides operator compatibility checks against the TOSA specification. Please note that TOSA is currently only available for x86 architecture.

Please, install it into the same environment as MLIA using this command:

mlia-backend install tosa-checker

Additional resources:

Vela

The Vela backend provides performance metrics for Ethos-U based systems. Please, install it into the same environment as MLIA using this command:

mlia-backend install vela

Additional resources:

About

The ML Inference Advisor (MLIA) helps AI developers design and optimize neural network models for efficient inference on Arm® targets. MLIA provides insights on how the ML model will perform on Arm early in the model development cycle and advice operator compatibility, performance analysis and model optimization.

Resources

Readme

Contributing

Contributing

Security policy

Security policy
Activity
Custom properties

Stars

0 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases

22 tags

Packages

No packages published

Contributors 9

Languages