Parameters of Neural Networks

In this set of parameters, aspects related to the architecture of neural networks that will learn from the dataset examples in training are controlled.

Each type of neural network has its own architecture, weight and performance parameters. Eyeflow.AI is an extensible platform that allows you to work with the most diverse architectures. However, our experience of several projects has already taught us about various architectures that work well in production, and it is this experience that we seek to bring to the platform, and thus simplify the lives of users.

In this Beta phase we have 2 main components that we use to solve all the problems that we have encountered.

Neural Network Parameters

Parameters for Neural Network

Classification Parameters

Classification Specific Neural Network Parameters

Parameter	Values	Default	Description
Component	choice [‘class_cnn’]	class_cnn	The DNN component for train model
Neural Network Depth	int [1 - 10]	3	Depth (num layers) of Neural Network
Neural Network Width	int [5 - 128]	20	Wide (num features) of Neural Network
Preprocess Mode	choice [‘caffe’, ‘tf’]	caffe	Function for image normalize
Loss Function	choice [‘categorical_crossentropy’, ‘binary_crossentropy’]	categorical_crossentropy	Loss function for use in training
Metrics Functions	Array of string	[‘categorical_accuracy’]	Metrics functions for use in evaluation
Optimizer Function	choice [‘adam’]	adam	Optimizer function for use in training

ObjectDetection Parameters

ObjectDetection Specific Neural Network Parameters

Parameter	Values	Default	Description
Component	choice [‘objdet’]	objdet	The DNN component for train model
Neural Network Width	int [5 - 128]	20	Wide (num features) of Neural Network
Neural Network backbone	choice [‘vgg7’, ‘vgg16’, ‘vgg19’, ‘resnet18’, ‘resnet34’, ‘resnet50’, ‘resnet101’, ‘resnet152’, ‘mobilenet128’, ‘mobilenet160’, ‘mobilenet192’, ‘mobilenet224’, ‘densenet121’, ‘densenet169’, ‘densenet201’]	vgg7	Backbone architecture to Neural Network
Preprocess Mode	choice [‘caffe’, ‘tf’]	caffe	Function for image normalize
IoU negative overlap	number [0.05 - 1.0]	0.3	Value for minimum overlap of negative boxes
IoU positive overlap	number [0.05 - 1.0]	0.45	Value for minimum overlap of positive boxes

Anchor parmeters

Parameters for boxes anchors

Parameter	Values	Default	Description
Boxes sizes	Array of integer	[12, 24, 48, 96, 192]	Size of boxes in each layer
Boxes strides	Array of integer	[8, 16, 32, 64, 128]	Strides of boxes in each layer
Boxes ratios	Array of number	[0.5, 1, 2]	Ratios (height / width) of candidate boxes
Boxes scales	Array of number	[1, 1.2, 1.5]	Scales of candidate boxes

Where should I go now?

• Training Parameters
• Parameters of Neural Networks
• Data Expansion Parameters