CS-231n Winter 2016 ( Convolutional Neural Networks for Visual Recognition )

Have a look at Course-Description

Cs-231n encountered its second series in 2016 taught by Prof Lei Fei Fei and Andrej Karpathy at Stanford University.

Lecture 1 : Data-driven approach, kNN, Linear Classification

1. Python

First of all, understanding the process of building neural networks requires a basic understanding of Python. I have compiled some important fundamentals in Python_Basics.py to get started.

Useful Links to get started :

• Python-Numpy Tutorial of CS-231N
• Python-Lists
• Python-Dictionaries
• Python-Sets
• Python-Tuples
• Numpy

2. Image Classification- A core task in Computer Vision

An image is composed of a 2d matrix of numbers also called pixels . For a colored picture we have 3 layers of Red, Green, Blue 2d matrices stacked together one by one.

What is Data Driven Approach ?

1. Collect a dataset of images and labels.
2. Use Machine Learning to train an image classifier.
3. Evaluate the classifier on a withheld set of images.

Lets start :

Classifier 1: K Nearest Neighbor

Dataset: CIFAR 10, Labels : 10, Training Images : 50000, Test Images : 10000

• Approach 1 : Manhatten Distance

  1.Import Dependencies

   import tensorflow as tf
   from tensorflow.keras.datasets import cifar10
   import matplotlib.pyplot as plt
   import numpy as np
  

  2.Load Data

   (x_train,y_train),(x_test,y_test)=cifar10.load_data()
  

  3.Finding 10 closest images

   a=np.array(x_train.reshape(50000,3072)) #Reshaping training data
   
   b=a-np.array(x_test[1].reshape(1,3072)) # Calculating Manhattan Distance from x_test[1](SHIP) with training data
   
   c=np.sum(b,axis=1) # Row wise summation
   
   index=np.argpartition(c,10) 
   s=c[index[:10]] # Finding the 10 closest images
   
   fig=plt.figure(figsize=(60,60))
  
   fig.add_subplot(2,12,1)
   plt.imshow(x_test[1]) # TESI IMAGE
   
   for i in range(1,11):
      fig.add_subplot(2,12,i+1)
      d=np.where(c==s[i-1])
      plt.imshow(x_train[d[0]][0])
  

  

  Result

  Manhattan distance performs poorly on getting similar images as many images are wrongly classified here.

  Check my CIFAR-10-Manhatten.py code for detailed implementation with some more examples

Lecture 2 : Linear Classification 2, Optimization