Convolutional Recurrent Networks

This module belongs to the " Computer Vision" category .

Description

The Convolutional Recurrent Network (CRNN or ConVNet) module is used to initiate a RNN estimator based on SmartPredict library.

This is a particularly popular function for treating visual data.

Parameters

As parameters, we can input shape ourselves right from the beginning. Then , we can adjust the optimization parameters and finally set the features selector.

Optimization parameters

The CRNN parameters are as follows:

• Learning rate

• Learning rate Reduction Factor

• Beta 1 and 2

• Gradient Clipping by Norm/Value

• Epsilon

• Momentum

• Rho

Optimizers

As optimizers :

• adam

• rmsprop

• adagrad

• adamax

• sgd (stochastic gradient descent)

The CRNN.

Features extractor

The CRNN module parameters allow to implement different kinds of features extractor:

• fcn

• darknet

• mobilenet_v2

• resnet50_v2

• resnet101_v2

• resnet152_v2

• nasnet

• xception

We can select one from the various features extractors.

The output decoder can be ctc or simple and imagenet is available as a features extractor weight.