Generating a convolutional neural network using the Pytorch library for image classification

This project features the necessary steps to generate a complete CNN for image classification in Python using the Pytorch library. The Datasets used for this project are the ones of CIFAR 10. CIFAR-10 consists in 60k (32x32) RGB images, labeled with an integer from 0 to 9, corresponding to 10 classes mentioned as following.

The classifications of the images are determined as follows:

1. plane
2. car
3. bird
4. cat
5. deer
6. dog
7. frog
8. horse
9. ship
10. truck

Neural Network model Architecture

The Neural Network model is composed by 6 layers, from which the first 3 are convolutional and the rest are linear layers. The architecture is found as follows:

3 x 32 x 32 (input)

    | kernel size = (3,3); padding = 1; stride = 1; output channels = 16;
    V  operation: Convolution; activation: ReLu;

16 x 32 x 32

    | kernel size = (2,2); stride = 2; 
    V operation: Max Pooling

16 x 15 x 15

    | kernel size = (3,3); padding = 1; stride = 1; output channels = 32; 
    V operation: Convolution; activation: ReLu;

32 x 15 x 15

    | kernel size = (2,2); stride = 1;
    V operation: Max Pooling;

32 x 8 x 8

    | kernel size = (3,3); padding = 1; stride = 1; output channels = 64;
    V operation: Convolution; activation ReLu;

64 x 8 x 8

    | kernel size = (2,2); stride = 1;
    V operation: Max Pooling;

64 x 4 x 4

    | operation: Flatten; output: 1-dimensional tensor with size (1024);
    V

500

    | operation: Linear; activation: ReLu;
    V

128

    | operation: Linear; activation: ReLu;
    V

10

    | operation: Linear; activation: logaritmic softmax;
    V

OUTPUT CLASSIFICATION

CLASSIFICATION RESULTS

The resulting classification is represented as the probability of the image being each of the forementioned classes. Each of the class labels has a probability value presented as 0 < p(x) < 1 for x in each label. Once all of the probabilities p(x) have been calculated in the forementioned CNN, a comparison is performed between each of the probabilities of the classes, and the highest probability is the resulting class of the image.

Datasets Normalization

The normalization vectors in the main fuction are calculated following the procedure as follows:

1. Get all the images in the training portion of the dataset, and stack them in a (k, j, i, n) tensor.

In the resulting tensor the representations are determined as follows: k: color channels of each image; j: the height of each image; i: the width of each image; n: the number of color images in the dataset;

2. Once this 4-d tensor has been calculated, flatten all of the dimensions excluding the first in order to get a 2-d tensor with dimensions (k, j*i*n), hereby known as R tensor, with dimensions (k, m).

3. Calculate the mean and vector std vectors by operating on the dimension m of the resulting R tensor.