Deep Learning accuracy validation framework¶
The Accuracy Checker is an extensible, flexible and configurable Deep Learning accuracy validation framework. The tool has a modular structure and allows to reproduce validation pipeline and collect aggregated quality indicators for popular datasets both for networks in source frameworks and in the OpenVINO™ supported formats.
Install prerequisites first:
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 is tested only via virtual environment.
Install the virtual environment:
python3 -m pip install virtualenv python3 -m virtualenv -p `which python3` <directory_for_environment>
Activate the virtual environment:
Virtual environment can be deactivated using the following command:
The next step is installing backend frameworks for Accuracy Checker.
To evaluate some models, you need to install the required frameworks. Accuracy Checker supports the following frameworks:
You can use any of them or several at a time. For correct work, Accuracy Checker requires at least one. You can 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 -m pip install .
Accuracy Checker is a modular tool and have some task-specific dependencies, all specific required modules can be found in
requirements-extra.in file. Standard installation procedure includes only basic part, in order to obtain extra modules you can execute following command:
python3 -m pip install .[extra]
When previous version of the tool is already installed in the environment, in some cases, it can broke the new installation. If you get a directory/file not found error, try manually removing the previous tool version from your environment or install the tool using following command in Accuracy Checker directory instead of setup.py install:
python3 -m pip install --upgrade --force-reinstall .
accuracy_checkcommand failed with following error:
from .cv2 import \* ImportError: libGL.so.1: cannot open shared object file: No such file or directory
try to uninstall
opencv-python-headlesspackage. More detils about the error and approaches how to fix can be found here
Running the Tool inside IDE for Development Purposes¶
Accuracy Checker tool has an entry point for running in CLI, however, the majority of popular code editors or integrated development environments (IDEs) expect scripts as the starting point of application. Sometimes it can be useful to run the tool as a script for debugging or enabling new models. To use Accuracy Checker inside the IDE, you need to create a script in accuracy_checker root directory, for example,
<open_model_zoo>/tools/accuracy_checker/main.py, with the following code:
from openvino.tools.accuracy_checker.main import main if __name__ == '__main__': main()
Now, you can use this script for running Accuracy Checker in IDE.
You may test your installation and get familiar with Accuracy Checker by running a sample.
Each Open Model Zoo model can be evaluated using a configuration file. To learn more, refer to How to use predefined configuration files guide.
Once you installed accuracy checker, you can evaluate your configurations using:
accuracy_check -c path/to/configuration_file -m /path/to/models -s /path/to/source/data -a /path/to/annotation
-h, --help to get the full list of command-line options. Some arguments are described below:
-c, --configpath to configuration file.
-m, --modelsspecifies directory in which models and weights declared in config file will be searched. You can also 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, --sourcespecifies directory in which input images will be searched.
-a, --annotationsspecifies directory in which annotation and meta files will be searched.
-d, --definitionspath to the global configuration file.
-e, --extensionsdirectory with InferenceEngine extensions.
directory to store Model Optimizer converted models (used for DLSDK launcher only). --tf, target_framework
framework for infer. --td, target_devices
devices for infer. You can specify several devices using space as a delimiter. -async_mode
allows run the tool in async mode if launcher supports it. -num_requests
number requests for async execution. Allows override provided in config info. Default isAUTO
directory with additional models attributes. -subsample_size
dataset subsample size. -shuffle
allows shuffle annotation during creation a subset if subsample_size argument is provided. Default isTrue
enables intermediate metrics results printing. Default isFalse
number of iterations for updated metrics result printing ifintermediate_metrics_results` flag enabled. Default is 1000.
You are also able to replace some command-line arguments with the environment variables for path prefixing. Supported list of variables includes:
DEFINITIONS_FILE- equivalent of
DATA_DIR- equivalent of
MODELS_DIR- equivalent of
EXTENSIONS- equivalent of
ANNOTATIONS_DIR- equivalent of
MODEL_ATTRIBUTES_DIR- equivalent of
There is a config file, which declares the validation process. Every validated model has to have its entry in the
models list with distinct
name and other properties described below.
There is also a definitions file, which declares global options shared across all models. Config file has priority over definitions file.
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 at
<omz_dir>/data/dataset_definitions.yml. 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 (
Launcher is a description of how your model should be executed. Each launcher configuration starts with setting
framework name. Currently caffe, dlsdk, mxnet, tf, tf2, tf_lite, opencv, onnx_runtime, pytorch, paddlepaddle supported. Launcher description can have differences.
Dataset entry describes the 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 (
<omz_dir>/data/dataset_definitions.yml). 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. Find more detailed information about dataset preparation 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.
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
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 a
type field with other options specific to the type. If you do not provide any other option, it will be picked from the definitions file.
You can use the following instructions:
You may optionally provide
reference field for metric, if you want the calculated metric tested against a specific value (reported in canonical paper).
Some metrics support providing vector results, for example, mAP is able to return average precision for each detection class. You can change view mode for metric results using
presenter (for example,
metrics: - type: accuracy top_k: 5 reference: 86.43 threshold: 0.005
Testing New Models¶
Typical workflow for testing a new model includes:
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.
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).
Reproduce preprocessing, metrics and postprocessing from canonical paper.
Create entry in config file and execute.
Standard Accuracy Checker validation pipeline: Annotation Reading -> Data Reading -> Preprocessing -> Inference -> Postprocessing -> Metrics. In some cases, this validation pipeline can be unsuitable, for example, when you have a sequence of models. You can customize validation pipeline using your own evaluator. Find more details about custom evaluations in the related section.