Deep Learning accuracy validation framework

Installation

TIP: You also can work with the Accuracy Checker inside the OpenVINO™ Deep Learning Workbench (DL Workbench). DL Workbench is a platform built upon OpenVINO™ and provides a web-based graphical environment that enables you to optimize, fine-tune, analyze, visualize, and compare performance of deep learning models on various Intel® architecture configurations. In the DL Workbench, you can use most of OpenVINO™ toolkit components.
Proceed to an easy installation from Docker to get started.

Prerequisites

Install prerequisites first:

1. Python

accuracy checker uses Python 3. Install it first:

sudo apt-get install python3 python3-dev python3-setuptools python3-pip

Python setuptools and python package manager (pip) install packages into system directory by default. Installation of accuracy checker tested only via virtual environment.

In order to use virtual environment you should install it first:

python3 -m pip install virtualenv
python3 -m virtualenv -p `which python3` <directory_for_environment>

Before starting to work inside virtual environment, it should be activated:

source <directory_for_environment>/bin/activate

Virtual environment can be deactivated using command

deactivate

2. Frameworks

The next step is installing backend frameworks for Accuracy Checker.

In order to evaluate some models required frameworks have to be installed. Accuracy-Checker supports these frameworks:

You can use any of them or several at a time. For correct work, Accuracy Checker requires at least one. You are able postpone installation of other frameworks and install them when they will be necessary.

Install accuracy checker

If all prerequisite are installed, then you are ready to install accuracy checker:

python3 setup.py install

Accuracy Checker is modular tool and have some task-specific dependencies, all specific required modules can be found in requirements.in file. You can install only core part of the tool without additional dependencies and manage them by your-self using following command instead of standard installation:

python setup.py install_core

Troubleshooting during installation

When previous version of the tool already installed in the environment, in some cases, it can broke new installation. If you see the error about directory/file not found, please try remove manually old tool version from your environment or install the tool with following command (in accuracy checker directory, instead of setup.py install):

pip install --upgrade --force-reinstall .

Usage

You may test your installation and get familiar with accuracy checker by running sample.

Each Open Model Zoo model can be evaluated using a configuration file. Please refer to How to use predefined configuration files guide.

Once you installed accuracy checker you can evaluate your configurations with:

accuracy_check -c path/to/configuration_file -m /path/to/models -s /path/to/source/data -a /path/to/annotation

You may refer to -h, --help to full list of command line options. Some arguments are:

  • -c, --config path to configuration file.
  • -m, --models specifies directory in which models and weights declared in config file will be searched. You also can specify space separated list of directories if you want to run the same configuration several times with models located in different directories or if you have the pipeline with several models.
  • -s, --source specifies directory in which input images will be searched.
  • -a, --annotations specifies directory in which annotation and meta files will be searched.
  • -d, --definitions path to the global configuration file.
  • -e, --extensions directory with InferenceEngine extensions.
  • -b, --bitstreams directory with bitstream (for Inference Engine with fpga plugin).
  • ‘-C, ’–converted_modelsdirectory to store Model Optimizer converted models (used for DLSDK launcher only). --tf, –target_frameworkframework for infer. --td, –target_devicesdevices for infer. You can specify several devices using space as a delimiter. -–async_modeallows run the tool in async mode if launcher support it. -–num_requestsnumber requests for async execution. Allows override provided in config info. Default isAUTO -–model_attributesdirectory with additional models attributes. -–subsample_sizedataset subsample size. -–shuffleallow shuffle annotation during creation a subset if subsample_size argument is provided. Default isTrue`.

You are also able to replace some command line arguments with environment variables for path prefixing. Supported following list of variables:

  • DATA_DIR - equivalent of -s, --source.
  • MODELS_DIR - equivalent of -m, --models.
  • EXTENSIONS - equivalent of -e, --extensions.
  • ANNOTATIONS_DIR - equivalent of -a, --annotations.
  • BITSTREAMS_DIR - equivalent of -b, --bitstreams.
  • MODEL_ATTRIBUTES_DIR - equivalent of --model_attributes.

Configuration

There is config file which declares validation process. Every validated model has to have its entry in models list with distinct name and other properties described below.

There is also definitions file, which declares global options shared across all models. Config file has priority over definitions file.

example:

models:
- name: model_name
launchers:
- framework: caffe
model: bvlc_alexnet.prototxt
weights: bvlc_alexnet.caffemodel
adapter: classification
batch: 128
datasets:
- name: dataset_name

Optionally you can use global configuration. It can be useful for avoiding duplication if you have several models which should be run on the same dataset. Example of global definitions file can be found here. Global definitions will be merged with evaluation config in the runtime by dataset name. Parameters of global configuration can be overwritten by local config (e.g. if in definitions specified resize with destination size 224 and in the local config used resize with size 227, the value in config - 227 will be used as resize parameter) You can use field global_definitions for specifying path to global definitions directly in the model config or via command line arguments (-d, --definitions).

Launchers

Launcher is a description of how your model should be executed. Each launcher configuration starts with setting framework name. Currently caffe, dlsdk, mxnet, tf, tf_lite, opencv, onnx_runtime supported. Launcher description can have differences. Please view:

Datasets

Dataset entry describes data on which model should be evaluated, all required preprocessing and postprocessing/filtering steps, and metrics that will be used for evaluation.

If your dataset data is a well-known competition problem (COCO, Pascal VOC, and others) and/or can be potentially reused for other models it is reasonable to declare it in some global configuration file (definition file). This way in your local configuration file you can provide only name and all required steps will be picked from global one. To pass path to this global configuration use --definition argument of CLI.

If you want to evaluate models using prepared config files and well-known datasets, you need to organize folders with validation datasets in a certain way. More detailed information about dataset preparation you can find in Dataset Preparation Guide.

Each dataset must have:

  • name - unique identifier of your model/topology.
  • data_source: path to directory where input data is stored.
  • metrics: list of metrics that should be computed.

And optionally:

  • preprocessing: list of preprocessing steps applied to input data. If you want calculated metrics to match reported, you must reproduce preprocessing from canonical paper of your topology or ask topology author about required steps.
  • postprocessing: list of postprocessing steps.
  • reader: approach for data reading. Default reader is opencv_imread.
  • segmentation_masks_source - path to directory where gt masks for semantic segmentation task stored.

Also it must contain data related to annotation. You can convert annotation in-place using:

  • annotation_conversion: parameters for annotation conversion

or use existing annotation file and dataset meta:

  • annotation - path to annotation file, you must convert annotation to representation of dataset problem first, you may choose one of the converters from annotation-converters if there is already converter for your dataset or write your own.
  • dataset_meta: path to metadata file (generated by converter). More detailed information about annotation conversion you can find in Annotation Conversion Guide.

example of dataset definition:

- name: dataset_name
annotation: annotation.pickle
data_source: images_folder
preprocessing:
- type: resize
dst_width: 256
dst_height: 256
- type: normalization
mean: imagenet
- type: crop
dst_width: 227
dst_height: 227
metrics:
- type: accuracy

Preprocessing, Metrics, Postprocessing

Each entry of preprocessing, metrics, postprocessing must have type field, other options are specific to type. If you do not provide any other option, then it will be picked from definitions file.

You can find useful following instructions:

You may optionally provide reference field for metric, if you want calculated metric tested against specific value (i.e. reported in canonical paper).

Some metrics support providing vector results ( e. g. mAP is able to return average precision for each detection class). You can change view mode for metric results using presenter (e.g. print_vector, print_scalar).

example:

metrics:
- type: accuracy
top_k: 5
reference: 86.43
threshold: 0.005

Testing new models

Typical workflow for testing new model include:

  1. Convert annotation of your dataset. Use one of the converters from annotation-converters, or write your own if there is no converter for your dataset. You can find detailed instruction how to use converters in Annotation Conversion Guide.
  2. Choose one of adapters or write your own. Adapter converts raw output produced by framework to high level problem specific representation (e.g. ClassificationPrediction, DetectionPrediction, etc).
  3. Reproduce preprocessing, metrics and postprocessing from canonical paper.
  4. Create entry in config file and execute.

Customizing Evaluation

Standard Accuracy Checker validation pipeline: Annotation Reading -> Data Reading -> Preprocessing -> Inference -> Postprocessing -> Metrics. In some cases it can be unsuitable (e.g. if you have sequence of models). You are able to customize validation pipeline using own evaluator. More details about custom evaluations can be found in the related section.