Assignment_2

Programming Assignment 2
Cloth Classification using Deep Neural Networks in numpy
Imports:
Use only numpy and a few other basic python libraries. Do not use any automatic differentiation libraries
like Tensorflow or Pytorch.
Task:
Build a neural network with 1 hidden layer and perform the cloth classification task
Neural network specifics:
1. No of hidden layers : 1
2. Non-linearity in all layers : relu
3. Forward pass equation : y_ = softmax((W2(relu(W1x + b1)) + b2) wherein x is your input item and
y_ is the vector of same size as that of no of classes. It gives the probability that x belongs to a
particular class.
4. Use softmax at the output of the final layer to calculate probability of the item belonging to each
class. (Be careful and check that your softmax function calculates correct values.)
5. Loss function: categorical cross entropy
6. Regularization: Use L2 regularization to limit the capacity of your model to prevent overfitting.
Data: https://goo.gl/dYFTP7
● train_images: 60000 examples of 28x28 grayscale image. The data is of size 60000 x 784. Each
pixel has a single intensity value which is an integer between 0 and 255.
● train_labels: 60000 labels from 10 classes for the images in given in train_images. Class details
are mentioned below.
● test_images: 10000 examples of 28x28 grayscale image. The data is of size 10000 x 784. Each
pixel has a single intensity value which is an integer between 0 and 255.
● test_labels: 10000 labels from 10 classes for the images in given in test_images. Class details
are mentioned below.
Labels
Each training and test example is assigned to one of the following labels:
● 0 T-shirt/top
● 1 Trouser
● 2 Pullover
● 3 Dress
● 4 Coat
● 5 Sandal
● 6 Shirt
● 7 Sneaker
● 8 Bag
● 9 Ankle boot
Important Note: Please make sure that the data for training and testing is placed in '../data/'
directory i.e. present in the parent directory of python files being created. You don't have to
submit this data files in your final submission. Also do not change the names of any files.
Code Outline:
Code outline for all files in implementation section is provided here:
https://github.com/cs60010/Tutorials/tree/master/code
Load Data:
Function for reading data from zip files have been implemented in the code outline that is provided. It will
work if all the zip files for train and test data are in the data directory.
Implementation
Create the following files for writing your code:
1. data_loader.py : Use this file to load all datasets and create mini batches
2. module.py : Create a neural network class here. Initialize all weights in the __init__() constructor.
Create a member function forward() where you do a forward pass of the neural network. Create
a member function backward() where you compute the loss and do a backward
pass(backpropagation) of the neural network and update all weights.
3. train.py : Implement a simple minibatch SGD loop and train your neural network. Finally it should
print the final accuracy after training.
Training and testing the network: python train.py should train the neural network, print out all losses on
train data and the final accuracy.
Store the final accuracy in a file ‘accuracy.txt’. Submit this along with your code. We should get a similar
accuracy as reported in ‘accuracy.txt’ when we run your code. Make sure you use a fixed random seed in
your code.
Reading Resources:
1. For calculating vectorized gradients during backprop:
http://www.psi.toronto.edu/~andrew/papers/matrix_calculus_for_learning.pdf
2. A more detailed matrix calculus tut: http://cs231n.stanford.edu/handouts/derivatives.pdf
Submission Instructions
● You have to use only numpy and basic python functions and no other imports are allowed.
● Implement the forward and backward passes of your network using vectorised numpy operations.
● Please make sure that the path for the data is same as demonstrated in the command to read
labels and images in the code outline.
● Please submit the completed compressed file named as 'Assignment2_<Roll No>.zip’
containing the files mentioned in implementation section. Please don't submit any data files.
Bonus points:
Make some interesting observations or provide some insights about the model that you have built.

Assignment_2

Uploaded by

Assignment_2

Uploaded by

Programming Assignment 2

Cloth Classification using Deep Neural Networks in numpy

You might also like