Career skills to jumpstart your future.

LightSpeed Intro to ConvNets

Handcrafted by On

LightSpeed intro to ConvNets

________________________________________________
Unless you're experienced or have read the LightSpeed intro to DL, please see here before reading this.
---------------------------------------------------------------------------------
In this post, you'll learn about-
  1. What are Convolutions?
  2. Why Convolutions work great (optional)
  3. Keras and TensorFlow (compulsory)
  4. How to make a Neural Network that can classify pictures depending on objects!

What are ConvNets?

Now, what are ConvNets? They are Neural Networks having Convolutions. Okay, so what are Convolutions? 
Convolutions, informally, are mathematical functions.
Now, to convolve the above matrices, we take the second smaller matrix, then we imprint it on the first possible 2 by 2 chunk of the other as shown below.

 So, this will be the first value in the output matrix-
= (4*1) + (2*0) + (0*0) + (3*1) = 7

Similarly, you may be able to calculate that the next output(Hint: Go one step right) shall also be 7, and the next one will be 12 and then, going one level down, we will get 6,12, and 7 as our values and so on.

Main points-

  1. We moved 1 block at a time, sideways and downwards. Thus, we used a stride of 1.
  2. We did not increase the size of the first matrix. Thus, we used valid padding.
  3. Also, we used a 2-by-2 filter (small matrix). Thus, we used the filter size of 2.
Thus-
s = 1
p = 0
f = 2

Why are ConvNets so good? (Optional)

There are two main widely accepted reasons why these ConvNets or Convolutional Neural Networks, work as well as they do.

  1. Parameter Sharing - For 1280 * 720 pixels picture (HD), with the usual three channels (Red Green Blue) a one-layer shallow Neural Network, will have the number of parameters running deep into the millions, with a single layer having exactly 2,746,800 parameters needed for a direct Fully Connected layer (excluding bias). A ConvNet, however, can have as less as 75 * x parameters (f = 5, excluding bias, only one convolution), x may ideally be about 32 or 64, in which case there will be 2400 or 4800 parameters, which is still exponentially lesser than a fully connected layer.
  2. The sparsity of connections- Each output value depends on very few input values. 
Done!

Keras and Tensorflow 

TensorFlow is a relatively high-level machine learning programming framework made by Google. The later versions of TensorFlow included a library that was previously individually used. This was an even-higher-level deep learning programming framework- Keras. TensorFlow v1 and TensorFlow v2 had significant changes, which are beyond the scope of this blog post. Yay! We're done with the second part!

How to make a NN that classifies pictures

Like last time, we will use the GPU available on Colaboratory.
  • Download the notebook here on GitHub
  • Please star the Git if you liked the code explanation.
  • Upload notebook on Google Colab, and Enable GPU Hardware Accelerator.
  • Understand the code first. Ever welcome to ask doubts. Run both cells.
  • Play around with the code. Tune the values.
  • Try to achieve the highest accuracy possible!!! Above 90% is an achievement, so please comment below.
  • You are now a CNN programmer!!! Do write Feedback on both the GitHub Repository and the Blog.
  • Tell the LightSpeedMaster bot for what you want to learn next!!

I hope you enjoyed it!!


    Hi! I Love AI, DL and ML

    Comments

    Post a Comment

    Popular posts from this blog

    LightSpeed Basic Animation for Beginners in Flash 8 | Part-2

    Handcrafted by On
    Image
    Basics in Animation for Beginners in Flash 8 Now that you know about the basics of Flash 8 and the two basic types of tweening, get to know about the few basic tools that you will need to make a basic Flash 8 movie. If you're new around here, check out the Part-1 of the Basics in Animation for Beginners in Flash 8 .    Check out our blog LightSpeedAC for more amazing content.  Before we get started, even if you're not new around here, I would recommend you to revisit the Part-1 to revise the basic constituents of Flash 8.  Let Us Get Started This blog post will focus on the main basic tools that you need to use to make a simple, yet good Flash movie/animation.  1. Pen Tool - The Pen Tool is used to draw precise paths, like straight lines and smooth curves. You can draw straight lines or curves with this tool, and later adjust the angle and length of straight-line segments and the slope of curves.  The steps to use the Pen Tool are:...
    Read more

    LightSpeed Basic Animation for Beginners in Flash 8 | Part-1

    Handcrafted by On
    Image
    Basics of Animation in Flash 8 for Beginners  This blog will focus on the introduction to Flash 8 for Beginners, uses of Flash 8 in this ever-growing technological world. This blog will also include basic things that you need to know as a beginner in Flash 8 and the basic types of animations in Flash 8. If you're reading this blog post, you must have Flash 8 installed. You can get your 30-day free trial on Adobe's website and need to later buy it. You will be given a serial code. What is Flash 8? Macromedia Flash is a program that is used in creating interactive, and animated media such as motion graphics, videos, presentations and websites. You can control an object’s actions and behaviors using the Behavior panel and the Timeline. For more advanced control you can use Actionscript, Flash's programming language that is very similar to Javascript. Flash 8 or Macromedia Flash 8 was developed by Macromedia. Macromedia was acquired by Adobe in December of 2005. Sinc...
    Read more

    A honest review of the Data and Deployment Specialization

    Handcrafted by On
    Image
      What is the TensorFlow: Data & Deployment Specialization? The TensorFlow: Data and Deployment specialization is deeplearning.ai' s new group of courses and is intended to teach you how to practically use your DL skills to make practical applications. It is co-taught by Andrew Ng and Laurence Moroney, both seasoned veterans in AI and DL. I will give a detailed review of the first course below. I am trying to keep the content coming while doing the course myself! 1. Browser-Based models with TensorFlow.jS This course is intended to teach you how to set up a browser-based model using TensorFlow's browser equivalent- TfjS.  It shows the necessary JavaScript required for training models. It must be noted that tfjs is not direct TensorFlow! It's a different version, and they have few differences, which are correctly taught in the course. If you are used to Andrew Ng lectures, this isn't like that. The course assumes that you know Deep Learning, so most talks...
    Read more

    LightSpeed intro to Deep Learning

    Handcrafted by On
    Image
    LightSpeed intro to DL In this post, you'll learn about- Basic necessities of DL theory Keras and TensorFlow (optional) How to make a Neural Network that can identify digits in a picture! Basic Necessities of Deep Learning Theory Now, what is Deep Learning ? It is, plainly the art of training Neural Networks . Okay, so what are Neural Networks? They are just another type of Machine Learning model, which apparently work like the brain. And now, two questions remain- 'What do they do?' and 'How do they work?' The answer to the first question is that they first predict some output for a problem using layers , with randomly initialized parameters. Then, after seeing how far they are from the actual truth, they change their settings to obtain lesser loss or distance from the truth.  Now for the second question, the answer is much less mystified and more mathematical. It computes a non-linear function (usually a ReLU of the (input if it's the fir...
    Read more