Convolution neural network equations

Hello, all

I hope you doing well.

I know in CNN training data such as (mnist, CIFAR_10 or CIFAR_100 or even ImageNet dataset) there are

feedforward process and then compute the error hence, using backpropagation to minimize that error by updating whole weights in all layers,

however, I need to know how it implements in code for both feedforward and backpropagation progress?

Is there any code to interpret these equations especially equations of updating of all weights?

Thank you for helping..

Answer 1

[Mahesh Taparia]

Hi

As you know about the feedforward and backpropagation process, for feedforward network we need to create the network architecture. You can refer to this documentation for creation of a sample network. For backpropagation, target is to reduce the loss by finding the optimum weights. In this case the weights are getting updated by the equation:

newWeights=previousWeights-learningRate*derivative of loss wrt weights

In documentation, the direct inbuilt functions are used for backpropagation and weight update. However, you can create the custom functions in MATLAB with the help of the respective equations but inbuilt functions are more optimized.

[Ali Najem]

Many thanks for your answer the most important thing to me is how to obtain all equations of updating weights in backpro. ? How can i find this inbulit function ?

[Mahesh Taparia]

For equations, you can refer to a standard book of Deep Learning where you will get a sample derivation/ explaination on that. With that, you can derive for your network.

You can use 'dlgradient' function to compute the automatic differentiation. You can refer this documentation of dlgradient. Also, you can use function like 'adamupdate' to update the weights. You can refer its documentation here.

For complete flow to train a network for a custom training loop, you can refer to this documentation. Link to train network using trainingOptions is shared in previous answer. Hope it will help!

[Ali Najem]

Thank you so much sir, it helped

now am trying to use inception v3 with CIFAR-10 dataset

am new in this matter, how can i proceed this?

I found many programs just to know how to train CIFAR-10 by using inception v3 model but didn't work !!

Thanks for your helping ..

[Mahesh Taparia]

Hi

Refer this link for CIFAR data classification. The network architecture is different. Replace the network by inception v3 using 'inceptionv3' function. Refer its documentation here. In this network, the number of classes are 1000, replace the layers with 10 nclasses. For this, use 'replaceLayers' function to replace the last layer with number of classes as 10. Refer this documentation for more information. Hope it will help!