Deep Learning has made remarkable progress over the past few years with quick transitions from discovery of new methods to their industrial implementation. While framework and libraries have made creating and working with deep architectures easy, quite less is known by practitioners about the internal states of the process. This post is an attempt to find out what composes a neural network and what a convolutional neural network sees in an input.
After training, we load the layer to visualise and pass a sample input via the input layer. The function then runs a session for the layer given the input and returns all the filters that comprise that layer.
This is done by using TensorFlow's session.run() function which returns all the filters when a layer is fed in as an object.
Sample Input:
The results for the sample input are the following visualisations which are plot using matplotlib.
Hidden Layer 1 :
Hidden Layer 2:
The number of plots correspond to the increased number of filters as we go deeper into the network.
The depth also describes how more finer details are sought by the filters as the depth increases. This can be seen in the representation between what HiddenLayer1 vs HiddenLayer2 sees as the filter shows how the input stimulates the filter.
The height of your accomplishments equal the depth of your convictions.
Sound is diverse. It's nature is to comprise of so many forms that solid rules cannot draw lines between them. I always have been an admirer of sound.
I have spent a few weeks implementing sound acquisition and processing in android and have come up with something to begin with.
Presenting:
Shh..Silence
An application that monitors sound level of an environment and plays a 'Shhhh...' when noise levels pass a limit.
You can download it here:
Google Play and the Google Play logo are trademarks of Google LLC.
For the end user, it is a harmless application but from an engineering point of view, the internals are a window to a ton of possibilities.
The application monitors sound in the following manner :
Acquire Microphone
Configure
Sample Rate
Mono/Stereo
Encoding format
Buffer size
Calculate average over the buffer size
Compare obtained value with threshold
Trigger when above threshold
The pipeline is simple when dealing only with the average. Here's a code snippet for step 3:
private void readAudioBuffer() {
try {
short[] buffer = new short[bufferSize];
int bufferReadResult = 1;
if (audio != null) {
bufferReadResult = audio.read(buffer, 0, bufferSize); //Audio Samples
double sumLevel = 0;
for (int i = 0; i < bufferReadResult; i++) {
sumLevel += buffer[i];
}
lastLevel = Math.abs((sumLevel / bufferReadResult));
}
} catch (Exception e) {
e.printStackTrace();
}
}
The most intriguing part of it is at bufferReadResult. It contains a sequence of numbers that depict the sound received by the microphone. Following this, it is a matter of requirement what needs to be done next. On extracting audio features like Mel Coefficients, MFCC etc, the application can be stretched to domains of Audio Classification, Speech Recognition, User Identification and Keyword Detection.
Implementing ML/DL on android has become easier than ever using TensorFlow with a light framework. The next step is to develop an application that uses TensorFlow for purpose of classifying sounds.
We have always sought of new ways to interact with computers. From typed commands to automatic speech recognition, the aim is to make it appear natural to us, as if not interacting with a computer but a Human.
Project AlphaUI
AlphaUI is a virtual menu interface that lets you interact naturally with the GUI displayed. It works by using a webcam to capture live frames and through Image Processing finds out where the user wants to point out in the given space.
The Program is written in C++ using OpenCV 3.1.0 Library and performs the following operations on each image from which relevant information is extracted.
For the project to be demonstrated, I have utilised my computer vision project: Automatic Face Recognition System. The AlphaUI interface is built on top of the Face Recognition System with a custom GUI giving integrity to both projects.The functional response of interface have been disabled for the demo. Any developer can define their own GUI for their system that require user interaction in the same way.
Screenshots of the system:
The AlphaUI interface
Ball tracked continuously by the system
Touchless Interaction with the interface
The system can be trained on any object of interest provided it is distinct in color (read HSV segmentation) . The training is done by repeatedly marking all over the object with the mouse pointer. This step has to be done only once in a lifetime or when you need to use a new marker. The values are saved in a text file to be reused in next run.
Disclaimer:
This project was done about an year ago but never saw daylight until now. What would you do with the possibilities of this project? Do comment and let me know.
Project Phase A Face Recognition system to be used for marking attendance in an organisation for a streamlined and centralized record of Employees or Members.
Phase includes the following stages:
A C++ program to detect and store faces. (Detection)
A Python Script to maintain and link available faces. (Linking)
A C++ program to fetch faces from a camera and compare them with available database. (Recognition)
A Python Script to update the record on Google Spreadsheets over a secure wireless connection. (Uploading)
All of this runs on a Raspberry Pi 3.
Phase is and has been my most ambitious project because of the way it works. And also because it is composed of three of my most relished domains: Embedded Linux, Machine Learning and Internet of Things.
Each Functional component took considerable time worth mentioning in this catalog however I also do acknowledge that a lot of hows and whys will still be skipped because they are really large in number.
Last but most important, Kudos to Stack Overflow and every developer that asked relevant questions for this project of mine to be completed even after so many dead ends I have encountered.
Project Phase:
My initial setup was Ubuntu 15.10 on an Intel core i5 laptop. Linux was a choice because I already planned to deploy the final project on a Raspberry Pi. This maintained familiarity with the development platform.
The entire project is too long to discuss every working bit in a single blog post so an extended summary is what this post is. What Phase Does: It Sees You, Remembers You, Recognizes You and Keeps a note of it on Google Drive !
Pretty Cool when you think about it.
Each task stated above uses a separate program linked together to work seamlessly.
It Sees You:
The Video is captured via the integrated webcam (when developing on ubuntu) and via a USB webcam (when run on Raspbian OS [Raspberry Pi]).This is made easy by always fetching the video from the default connected device.As Pi doesn't come with an inbuilt camera,default device is the USB webcam.Voila!
A C++ program linked with the OpenCV (build from source:make,make install) running a cascade Haar's Frontal Face classifier detects the faces in an image.The task of detection is the following two things:
Number of faces in the image
Segmenting ,Cropping and Resizing Faces
Detecting One Face and Saving to Database
Detecting One Face and Saving to Database
A Video Documenting the database creation is as shown:
It Remembers You:
The database of a face is created only when only a single face is detected by the classifier.This ensures that the database of a single individual contains images only of that individual.Before saving the faces are gray scaled and resized to 300 x 300 pixels.
A lot of guidance was received from OpenCV documentation.This includes the above folder structures to store faces.
Root Folder
Database of an Individual
Although the database of images is created successfully,For a program to actually "See" them,It is crucial that every image is properly documented along with the ID of the person they represent.This path creation and Labeling is done by a Python Script.
The Python script creates a record in .csv format which contains :
Full Path of the image
Label Corresponding to the Image
Since images are stored in a folder named after the id of the person, The name of the folder is infact the Label for our task.
The .csv file created looks like:
The .CSV file created is used by the next segment to fetch,load and train the Face Recognizer algorithm.
It Recognizes You The Task of Face Recognition is done by C++ Program written using OpenCV library.
The Face Recognition module is not native to the official source yet so the additional libraries are built using a new method I came up with as documented here.This method is more reliable than the conventional route.
The program fetches live feed from the default imaging device and processes it frame by frame.
The first task that the program performs is to train its Two classifiers on the training database and labels of images.The Two algorithms used are:
Eigenface is single class specific i.e. It finds the similarities between multiple images of same individual whereas FisherFace finds the differences between different individuals.The Collective and commonly agreed result of both these algorithms trained on the same set of images is used as a confirmation of a prediction.
The Haar's cascade is run to segment the faces which are the evaluated by the two algorithms and predictions are returned by both.The value of prediction is accurate 90% of the trials however it depends on the quality of images in the database.
Video Documenting Face Recognition:
Keeping a note on Google Drive: The task of connecting securely to google cloud is done by a python script. It uses the following package to do the task of accessing and updating attendance on google spreadsheet.
Oauth2client (Google Cloud Authentication Client)
Gspread (Google Spreadsheet API client)
PyOpenSSL (Python Open SSL package)
The Result of prediction (Roll No. or Unique ID) is given to the Cloud Connect Script as a command line argument. The script fetches the date of current day from the system.These two data elements are enough to mark a student as present.
The Logic here is always a tautology,
i.e. if a student 'A' arrives before the system ,he is marked as present for the current day.
if a student 'A' is absent, he never arrives for attendance before the system, hence he is not marked for that day thus stating him absent.
The Python Script connects to a google spreadsheet via valid security credentials and update the attendance onto it.The programming is done in such a way that it handles all the possible scenarios that can arise on the spreadsheet section. Few of the problem -> solution are:
Date Row not found -> Create row for Current Date. (When taking attendance on a new day)
Roll No not found -> Create column for Roll No. (When database is updated)
Date Row found, Roll No column not found -> Add Roll No column and write "Present" in current date row (Database updated during current day)
Date Row not found, Roll No. Column Found -> Add Date Row and write "Present" in current Roll No. column (Database Intact, Day changed )
The Data for the recognized individual is successfully updated in 3-4 seconds. This is slow compared to execution time of our Recognizer program however keeping in mind all the authorizations and Credential check every time, it for sure is a lead over other unsecured connections.
The Google spreadsheet is edited to give write access to our API token so that there is no conflict of permissions during write task.
Here is the video of Phase updating the attendance of a detected individual in real time:
The Pi Setup:
The Setup is done with a Dell VGA monitor using an HDMI to VGA converter to connect to Raspberry Pi. Additionally USB Webcam,Keyboard & Mouse are connected via USB port.The webcam lights are kept off because of high current surge of 6 LEDs. They barely make any improvements in lighting conditions anyway.
An 8GB Sandisk MicroSD card is loaded with NOOBS and Raspbian OS is installed.
OpenCV is built from source using my method for extra modules building as stated here.
CodeBlocks is installed from apt-get and code is copied to from the ubuntu system to Pi using a thumb drive.
Static path for database storage, database linking,fetching and cloud uploading are set to get around using command line arguments every time.The Detection stage still employs CL arguments to denote the person being databased.
The entire system is enclosed in a box as follows :
The LCD and the glowing Leds are part of a temperature monitoring system. It measures the temperature of the box internals to warn or ward off any heat damage. And that is an entirely different story for a later time.
The code for a dlib variant of the face detection and recognition project is available for access on my github here : https://github.com/sanjeev309/face-recognition-dlib-tensorflow-knn
You will need to modify the core code to suit your requirement for an attendance system.
Pull requests are welcome.
-----------------------------------------------------------------------------------------------------------------
Success is not final, failure is not fatal: it is the courage to continue that counts
Used :
Code::Blocks IDE
PyCharm IDE
Raspbian OS
Atmel Studio 7.0
This post is the second in the series of developing a neural net line follower. For the post covering dataset generation, go here
Once data is handy, the next objective in a Neural Net is to plan out its structure.
Neural nets are mostly used for uncertain and non-linear operations as is our task of pattern matching a.k.a Classification.
A neural net has mainly the following layers which map the input to the output
Input Layer
Hidden Layer(s)
Output Layer
The Number of Hidden Layer is a deterministic on the complexity of the task being performed.For example, A deep learning net might contain more hidden layers for a deeper defragmentation of data.However there is a trade off between Depth and Speed of execution, therefore for our case, a single hidden layer would be enough.
The Designed Network looks as depicted here:
8 : 1 mapping through a hidden layer of two nodes (and one bias unit)
Intuition says this will be sufficient for our task as a partition in half can easily tell which side of our sensor strip sees the line.
For activation function in our network, I have used tangent sigmoid function:
Tangent Sigmoid Function
For faster development time, the neural net designer tool in MATLAB is used.For the sake of understanding, an algorithmic view is as shown:
x = 1 X 8; {Input from 8 nodes}
W12 = 2 X 8; {Weight Matrix from Input layer to Hidden Layer}
Z1= x*(W12)T = (1X8 * 8X2)= 1X2; {Mapping from Input layer to Hidden layer}
A1= TanSig(Z1)= 1X2; {Activation of Hidden Layer}
W23 = 1 X 2; {Weight Matrix from Hidden layer to Output Layer}
Z2 =A1*(W23)T = (1X2 * 2X1)=1X1; {Mapping from Hidden layer to Output Layer}
A2=TanSig(Z2); {Activation of Output Node}
Y=A2; {Output}
The specifications are set up in the NNFit tool in MATLAB and data obtained from the previous post is used to train the network using Backpropagation Algorithm. After the training completes, the entire process is stored as a script using the prompt window.
By default the script generated contains a lot of redundant information which can be optimised on examination.
The entire script is reduced to the following:
function [y1] = NNLF(x1)
%NNLF neural network simulation function.
%
% Generated by Neural Network Toolbox function genFunction, 13-Aug-2016 14:40:30.
%
% [y1] = NNLF(x1) takes these arguments:
% x = Qx8 matrix, input #1
% and returns:
% y = Qx1 matrix, output #1
% where Q is the number of samples.
%#ok<*RPMT0>
%(c)Sanjeev Tripathi ( AlphaDataOne.blogspot.in )
% ===== NEURAL NETWORK CONSTANTS =====
% Layer 1
b1 = [-8.8516132798193108e-10;2.1615423176361062];
IW1_1 = [-30.312171052276302 -15.230142543653209 -7.5989117276441904 -3.8426480381828529 3.8426480396510354 7.5989117277872076 15.230142543113857 30.312171051352024;1.1787493338995503 1.1684723902794487 -0.30187584946551604 -1.2505266965306716 -0.85655951742083458 -0.61361689937359887 -0.51938433720151178 0.43601182390986715];
% Layer 2
b2 = -3.5519126834821509e-10;
LW2_1 = [1.0000000099939996 2.6043564908190074e-09];
% ===== SIMULATION ========
Q = size(x1,1); % samples
x1 = x1';
xp1=2*x1 -1;
xp1=cast(xp1,'double');
a1 = tansig_apply(b1 + IW1_1*xp1);
a2 = repmat(b2,1,Q) + LW2_1*a1;
y1=a2;
end
% Sigmoid Symmetric Transfer Function
function a = tansig_apply(n,~)
a = 2 ./ (1 + exp(-2*n)) - 1;
end
The Neural Net being run uses the pre optimized weights to map the input to the output with an accuracy of 100% (High Variance) as the dataset contained all the possibilities. A demo of our net for different inputs entered manually is as shown:
As stated earlier, -1 informs to move left +1 informs to move right ~0 informs to keep moving forward
The network of nodes can compute with accuracy any width,orientation or order of line with absolute accuracy. It can also detect multiple lines simultaneously as the network has been trained for all the possible inputs that it can encounter.
An Implementation on AVR Atmega32 microcontroller to be covered soon.
Stay Tuned for more.
Neural Networks are convenient when mapping a function which behaves non-linearly. However, for Training and Testing purpose , Data is the most important piece of the puzzle.
The Neural Network Line Follower to be designed uses 8 Line sensing elements where each will return
1 : Line Detected
0 : Line Not Detected
So the possible dataset is the combination of 8 binary units arranged in different pattern leading to a total of
28 = 256 Samples of data.
I wrote the following MATLAB script to create the data for training the Neural Network.
function [bin,out]=DataGen(m)
% Function to Generate Training Data set for training neural network of a
% line follower
%
% [bin,out]=DataGen(number of inputs)
%
% It is a good practice to have an even number of input units for the
% NN Network to be trained
n=(2^m)-1;
bin=decimalToBinaryVector(0:n);
[p,q]=size(bin);
out=zeros(p,1);
bLeft=bin(:,1:(q/2));
bRight=bin(:,((q/2)+1):end);
wL=bi2de(bLeft,'left-msb');
wR=bi2de(bRight,'right-msb');
for i=1:p
if wL(i)>wR(i)
out(i)=-1;
end
if wL(i)<wR(i)
out(i)=1;
end
if wL(i)==wR(i)
out(i)=0;
end
end
end
Here: bin holds the binary input sequence of 8 bits out holds the output corresponding to the input bin Such that:
if bin=[1 0 0 0 0 0 0 0] then out = -1 (Line Sensed on far left : Move Left)
if bin=[0 0 0 0 0 0 1 0] then out = 1 (Line sensed on right : Move Right)
if bin=[0 0 1 0 0 1 0 0] then out = 0 (Line sensed symetrically : Keep Moving Forward) For a Neural Network, 0 is pretty much a perfection so it allocates a really small value (~0) such that it can be considered as zero.
From the script above with 8 as a parameter via the following syntax:
[bin,out]=DataGen(8);
we get:
bin = 256 X 8 Input Data Matrix
out = 256 X 1 Output Data Matrix
In the screenshot of the varibles, Left Segment shows the input from the sensors while we have the expected output in the blue bounded box on right:
These Data elements are ready to be used as Training Parameters for our Neural Network.
Further Development to be covered by subsequent posts.