Wiki says that a neural network is a network or circuit of neurons, or in a modern sense, an artificial neural network, composed of artificial neurons or nodes.These artificial networks may be used for predictive modeling, adaptive control and applications where they can be trained via a dataset. Self-learning resulting from experience can occur within networks, which can derive conclusions from a complex and seemingly unrelated set of information.
Look at this picture.
I created this file to act as a starter template for students like me that want to implement a neural network and learn from it.
Most of the time I wanted to learn neural netwok, I came across a number of different parameters like weight initialization, # of neurons, decay rate, learning rate, batch size etc. And I always thought that there must be something or a course that will tell you all of these because most of the teachers online say that it's all about trial and error and it is about trial and error.
But finally I thought of creating a neural network on my own and came across some the best teachers in the world Joseph Redmon aka Pjreddie. He has open-sourced a number of lectures and they are a must to watch if you want to learn computer-vision . This the course link : Ancient Secrets of Computer Vision
I strongly recommend that you atleast see Lectures 12 and 13 to have a basic understanding of neural networks
All the methods related to neural network are implemented in src\classifier.c. Check all the mentioned functions in this file.
A neural network is made up of a number of layers or fully connected layers. Each layer has following entites:
- Input neurons
- Weights corresponding to each neuron
- Current weight updates while backpropogating
- Past weight updates to use with momentum
- Activation Function to add non-linearity in model
- Output neurons
An important part of machine learning, be it classifiers or neural networks, is the activation function you use. This has been implemented in void activate_matrix(matrix m, ACTIVATION a) and is used to modify m to be f(m) applied elemetwise all along the output matrix.
The different Activation functions that I have implemented are:
- LINEAR --> Just a linear activation
- LOGISTIC OR SIGMOID --> Not so good due to vanishing gradients
- RELU --> Rectified-linear-Unit
- LRELU --> Leaky-Relu
- SOFTMAX --> Mainly used before the output layer
As we are calculating the partial derivatives for the weights of our model we have to remember to factor in the gradient from our activation function at every layer. We will have the derivative of the loss with respect to the output of a layer (after activation) and we need the derivative of the loss with respect to the input (before activation). To do that we take each element in our delta (the partial derivative) and multiply it by the gradient of our activation function.
Normally, to calculate f'(x) we would need to remember what the input to the function, x, was. However, all of our activation functions have the nice property that we can calculate the gradient f'(x) only with the output f(x). This means we have to remember less stuff when running our model.
In my view you must differentiate all the functions yourself and then see if I have implemented it rightly or not. Check this readthedocs for detailed information Activation Functions Described.
After learning about the gradients check this function void gradient_matrix(matrix m, ACTIVATION a, matrix d)
for implementation details.
Now we can fill in how to forward-propagate information through a layer. First check out our layer struct:
typedef struct {
matrix in; // Saved input to a layer
matrix w; // Current weights for a layer
matrix dw; // Current weight updates
matrix v; // Past weight updates (for use with momentum)
matrix out; // Saved output from the layer
ACTIVATION activation; // Activation the layer uses
} layer;
During forward propagation we will do a few things. We'll multiply the input matrix by our weight matrix w. We'll apply the activation function activation. We'll also save the input and output matrices in the layer so that we can use them during backpropagation. But this is already done for you.
Check matrix forward_layer(layer *l, matrix in).
Using matrix operations we can batch-process data. Each row in the input in is a separate data point and each row in the returned matrix is the result of passing that data point through the layer.
Back Propogation is the most difficult to see through for me but as I implemented all the things became clear.
We have to backpropagate error through our layer. We will be given dL/dy, the derivative of the loss with respect to the output of the layer, y. The output y is given by y = f(xw) where x is the input, w is our weights for that layer, and f is the activation function. What we want to calculate is dL/dw so we can update our weights and dL/dx which is the backpropagated loss to the previous layer. Check matrix backward_layer(layer *l, matrix delta).
First we need to calculate dL/d(xw) using the gradient of our activation function. Recall:
dL/d(xw) = dL/dy * dy/d(xw)
= dL/dy * df(xw)/d(xw)
= dL/dy * f'(xw)
We use void gradient_matrix(matrix m, ACTIVATION a, matrix d) function to change delta from dL/dy to dL/d(xw)
Next we want to calculate the derivative with respect to our weights, dL/dw.
dL/dw = dL/d(xw) * d(xw)/dw
= dL/d(xw) * x
but remember to make the matrix dimensions work out right we acutally do the matrix operiation of xt * dL/d(xw) where xt is the transpose of the input matrix x.
In our layer we saved the input as l->in. Calculate xt using that and the matrix transpose function in our library, matrix transpose_matrix(matrix m). Then calculate dL/dw and save it into l->dw (free the old one first to not leak memory!). We'll use this later when updating our weights.
Next we want to calculate the derivative with respect to the input as well, dL/dx. Recall:
dL/dx = dL/d(xw) * d(xw)/dx
= dL/d(xw) * w
again, we have to make the matrix dimensions line up so it actually ends up being dL/d(xw) * wt where wt is the transpose of our weights, w. Calculate wt and then calculate dL/dx. This is the matrix we will return.
After we've performed a round of forward and backward propagation we want to update our weights. Check void update_layer(layer *l, double rate, double momentum, double decay) .
Remember that with momentum and decay our weight update rule is:
Δw_t = dL/dw_t - λw_t + mΔw_{t-1}
w_{t+1} = w_t + ηΔw_t
We'll be doing gradient ascent (the partial derivative component is positive) instead of descent because we'll flip the sign of our partial derivative when we calculate it at the output. We saved dL/dw_t as l->dw and by convention we'll use l->v to store the previous weight change Δw_{t-1}.
Calculate the current weight change as a weighted sum of l->dw, l->w, and l->v. The function matrix axpy_matrix(double a, matrix x, matrix y) will be useful here, it calculates the result of the operation ax+y where a is a scalar and x and y are matrices. Save the current weight change in l->v for next round (remember to free the current l->v first).
Finally, apply the weight change to your weights by adding a scaled amount of the change based on the learning rate.
Now since you have learned how to backpropogate the error. The main task still remains and that is to calculate the error at the last layer so we can propogate it in the first place. For that the way to do so is to find out the error using a error function.
I have implemented two error functions:
Cross Entropy Loss: Calculates the cross-entropy loss for the ground-truth labels y and our predictions p. Cross-entropy loss is just negative log-likelihood (we discussed log-likelihood in class for logistic regression) for multi-class classification. Since we added the negative sign it becomes a loss function that we want to minimize. Cross-entropy loss is given by:
L = Σ(-y_i log(p_i)) for a single data point, summing across the different classes. Cross-entropy loss is nice to use with our softmax activation because the partial derivatives are nice. If the output of our final layer uses a softmax: y = σ(wx) then: dL/d(wx) = (y - truth) See why f'(wx) becomes 1 Softmax with Cross Entropy Loss. During training, we'll just calculate dL/dy as (y - truth) and set the gradient of the softmax to be 1 everywhere to have the same effect, as discussed earlier. This avoids numerical instability with calculating the "true" dL/dy and dσ(x)/dx for cross-entropy loss and the softmax function independently and multiplying them together.
See matrix Last_Layer_Loss_Cross_Entropy(data b, matrix p) for dL/dy implementation.
Mean-squared Error: Calculates the mean squared error between the ground truth y and our predictions p. MSE is a loss most of you would be familiar with but for the notion of describing everything I am writing once more here.
MSE = (1/n)Σ(y - (p_i))^2
Here n denotes number of rows in a matrix or distinct data points. Now the question remains to use cross-entropy loss with MSE. Lets do that : So assume that in the first forward feed you have adjusted all your hidden layer outputs based on the fact that each neuron output will be f(wx). And now you want to back-propogate the error.
I have calculated the Backprop-Error dL/dy here in this picture. Please have a look.
Now since you got the maths behind writing the Gradient of MSE with respect to the output of the softmax output.
Lets see the function in action.
See matrix Last_Layer_Loss_Mean_Squared(data b, matrix p) in src/classifier.c for implementation.
To create a model that can do something, you have to create a number of hidden layers in the first place and for all those hidden layers you have to think of the best suited activation function and different hyperparameters that come along.
Lets see how to create a new layer:
As you can see in line 245 src/classifier.c.
-
We initialize the
l->inmatrix andl->outwith one neuron. -
We do the weight initialization quite effectively.
-
There a few methods to do so. One is known as Xavier Initialization. See it here [Xavier Initialization] (http://proceedings.mlr.press/v9/glorot10a/glorot10a.pdf), here I initialize the weights as a random matrix of size
input*outputwith values that lie in the range(-sqrt(2/input), sqrt(2/input)). And this works pretty well. -
I have also implemented normal distribution with some mean
mand variances. Check line 116 formatrix normal_random_matrix(int rows, int cols, double m, double s)and it works onbox-mulleralgorithm. Check wiki for that.
-
We create the past weights updates as l->v, current weight updates l->dw of same dimensions as that of l->w.
-
Final Step is to have a activation function for the layer. And we do this by setting
l->activation.
After every time we backpropogate and do weight updation we have to do forward propagation.
Check line 261 for that matrix forward_model(model m, matrix X).
Now this function does is call the forward_layer function for all the layers in the model and set the weights.
X = foward_layer(m.(layers+i), X) sets X that is the input matrix to the output matrix at each layer.
Once a foward pass is completed, we have to backpropogate the error through the model. Check line 273 in the src/classifier.c for more information on the above method. We send the matrix Loss_D (which represents loss of each layer) in backward sense through the backward_layer(layer* working_layer, matrix Loss_D) .
See line 357. The function declaration looks like this : void train_model(model m, data d, int batch, int iters, double rate, double momentum, double decay).
Now lets see what this function do.
- The model trains for
int itersnumber of times. - Creates a random data batch of batch size
batchfrom the whole data points. - Do a forward_propagation taking in (model, batch.X)
- Calculate the cross entropy loss for the last layer.
- Calculate the loss matrix and backpropogate it through the model.
- Update the model weight through
update_model(m, rate/batch, momentum, decay). See for SGD the rate changes to rate/batchsize.
After reading each and every bit of information that I have written here you are equipped with most of information needed to write something similar to this repo. I strongly feel that you should create your own Neural Network in C specially because you will be defining all the matrix methods your-self unlike python where you have numpy for help and also the execution will be slow of your own methods.
You don't have to code everything from scratch. I have build this on top of my existing repo. Check the repo Image_Processing_Starter_Packge
-
Clone this repository and make sure that you have Gcc installed. If you want to use any other C compiler, feel free to do that in the MakeFile that's presents in the root of this repo. See line 15 in
Makefile. Also make sure you havemakeinstalled. -
Fire
makefrom a Cli. This will create all the the objects and link them to each other. -
Finally create you own neural nets and play with them as whole of the project has a python wrapper. So you just have to define models in a
.pyfile. Example models are present in the neuralNetworkTrainer.py file.
- MNIST DATASET
To run your model you'll need the dataset. The training images can be found here and the test images are here, I've preprocessed them for you into PNG format. To get the data you can run:
wget https://pjreddie.com/media/files/mnist_train.tar.gz
wget https://pjreddie.com/media/files/mnist_test.tar.gz
tar xzf mnist_train.tar.gz
tar xzf mnist_test.tar.gz
We'll also need a list of the images in our training and test set. To do this you can run:
find train -name \*.png > mnist.train
find test -name \*.png > mnist.test
- CIFAR DATASET
We have to do a similar process as last time, getting the data and creating files to hold the paths. Run:
wget http://pjreddie.com/media/files/cifar.tgz
tar xzf cifar.tgz
find cifar/train -name \*.png > cifar.train
find cifar/test -name \*.png > cifar.test
Since fo now each layer is connected to the next layer in a fully connected way. I want to implement some other layers also like Convolution Layer, Max Pooling Layer, Dropout Layer etc.
-
If you have some doubt reagarding any line of code email me right away or create an issue. Email
-
If you would like to create custom layers as I have written above. Create a issue and discuss with me as I want this repo to be as clean as possible.
Joseph Redmon Whatever I have learnt about neural networks and stitched together here is a result of divine knowledge of Joseph Redmon and his assignment vision-hw4. I have copied most of his code but I have also added some-things myself like the Mse Error. And I thank [Pjreddie]((https://pjreddie.com) for that.
Shivam Jalotra - (https://github.com/jalotra)
- I thank Pjreddie from the bottom of my heart for creating such a greate course and wonderful assignments.
- I think that everybody who wants to learn Computer Vision or Neural Network in general should watch do this course Ancient Secrets of Computer Vision
- Finally I thank you for reading through all this.