Description
Neural Networks
• The purpose of HW4 is to learn how to
construct and train convolutional neural
networks
• With the use of high-level APIs for pytorch
and tensorflow
• Dataset: MNIST – all ten classes
(i.e., not restricted to two selected digits)
Part A: pytorch
• Use pytorch nn.Sequential container
(https://pytorch.org/docs/stable/nn.html)
to build a Convolutional Neural Network
(CNN) for 10-class MNIST problem.
• Use Convolutional and Pooling layers, and
explore using different parameters (e.g. # of
filters).
• Follow up the convolutional part of the network
with a standard feedforward dense layers
(nn.Linear + ReLU activation).
• Explore the effect of using Batch normalization.
Part B: tensorflow
• Use Functional API from tensorflow Keras
(https://www.tensorflow.org/guide/keras/fu
nctional) to build a CNN network similar to
what you have built in Part A
Returning the Assignment
n Solution code should be written by you and you
only (no web/book/friend/etc. code)
n Upload through Blackboard
n A report in PDF
n Description of the architecture of your networks
n Plots of the loss on the training set through epochs of training
n Accuracy of the final trained networks on the test set
n Code in python
