Vaico - Visual Artificial Intelligence

This package contains Abstract class for Machine Learning models with explicit model protocol. Instantiate AbcModel for create a new model and ensure better server compatibility. Also contain predicted geometries and model architectures

AbcModel methods

Implement required methods and set a dict of the default model parameters (Intended for the user to modify the model behaviour) in a variable _defaults.

All the keywords parameters are loaded as instance variables and also stored in _defaults variable

from MLcommon import AbcModel

class NewModel(AbcModel):
    _defaults = {
        'labels': None,
        'train_epsilon': 0.15
        'any_other': None
    }               
  

Zip architectures

For share a new architecture compress it in zip format under the folder python for AWS lambda compatibility. Install the requirements of each architecture

Models

Each model is loaded from two files:

• zip_path: path to architecture file (.zip)
• model_path: path to model file (.ml)

Loading models example for Cascade

Loading models example for generating a Cascade configuration:

conf['main_model']['model'] = load_zip_model(
    conf['main_model']['model_path'],
    conf['main_model']['zip_path'])
for roi, models in conf['sub_models'].items():
    for model_conf in models:
        model_conf['model'] = load_zip_model(model_conf['model_path'], model_conf['zip_path'])

Prediction strategies:

Cascade

Class for combine multiple prediction models based on the output of a main detector model. The regions output of the detector are passed on other nodes (Classification or detection) based some rules.

Cascade Prediction Strartegy

• Get predictions from a main detector
• Results are passed to specialized classifiers/detector

All the predictions are obtained from other nodes Classifiers are run in a concurrent requests Inputs are: - Detector node direction - Subregion to be cropped from detector result - Subregions dict (detector-output : classifier-node)

Example

conf: {
    'max_concur_req': 10,
    'main_model':{
        'model':AbcModel,
    }
     'sub_models':
        {
            'person':

                [
                    {
                        'model': AbcModel,
                        'weights': (0,0,1,1),
                        'conditions':['square_h']

                    },
                ],
            ...
        }
}

Use all for pass all the subregions in a classifier

Multi

General model that load a models configuration Run multiple models that are independednt one for another

Example of node configuration

conf={
    'models': [AbcModel1, AbcModel2, Cascade,...],
}

Roi

The roi inside the located class by the detector is defined by weights Each weight from 0-1 of the values: (xi, yi, w, h) Where 0 -> xi | 1 -> xi+W Example: "roi_weights": (0.1, 1, 0.5, 1), "roi_conditions": ["center_x", ],

• Conditions:
  • center_x: center x direction
  • center_y: center y direction
  • square_w: square to width (after weighted)
  • square_h: square to height (after weighted)

Geometry

Package for representing data as Objects with common geometries. Used as an interface for prediction elements in Computer Vision problems. Each predicted object is represented as an Object. Which have geometry, label, score and subobject

Implemented:

• BoundBox
• Point
• Circle
• Polygon
• Polyline

Importing

For use geometries:

from MLgeometry import Point
P = Point(1,2,z=13)

Create from dict:

from MLgeometry import creator
o_dict = {
    'geometry': 'boundbox',
    'label': 'myBB',
    'score': 1,
    'subobject': None
}

new_obj = creator.create_geometry(o_dict)

Methods

Geometry special methods: Geometry._asdict() and Geometry._fromdict(dict)

Architectures

Yolo v4 Lite

Yolov4 port for inference with tensorflow lite

Instalation of TF-lite

pip install https://dl.google.com/coral/python/tflite_runtime-2.1.0.post1-cp35-cp35m-linux_x86_64.whl

Zip

Compressed file for deploy on lambda layer. Include MLcommon module

License

All rights of this library belongs to Vaico Visión Artificial and no one may distribute, reproduce, or create derivative works.

Vaico