CS 523 Summer Semester 2, 2021: Problem Set 3
July 29, 2021
Release Date: Thursday, July 29, 2021 at 01:15 EST
Due Date: Thursday, August 12, 2021 at 23:59 EST
Deliverables: Your completed source code package in a .zip archive. Please remove any pycache directories. This will be submitted on Gradescope. Reminder: Please feel free to work together. If you collaborate with others, please write your team members names¡¯ in your submission.
Last time, we took the next step by implementing element-wise dependent activation functions. This is super nice, as now we have a decent package going. Now, we¡¯re going to implement two new architectures of neural networks.
In this assignment, we will be implementing a 2d convolutional layer as well as a vanilla RNN. I have already writen quite a bit of code, including some new layers that you may or may not find useful (which are already completed and functional). I have left it up to you to complete the backwards method for two layers. Please copy all of the files in this .zip archive into your existing ps2 solution to improve our package. The following files that need to be completed are:
1. nn/layers/conv2d.py (100pts)
2. nn/layers/rnn cells/vanilla rnn cell.py (100pts)
The files test conv2d.py, test vanilla rnn single output.py, and test vanilla rnn sequences.py are your testing scripts. Please go ahead and run them in order to test your code. If the progress bar completes and you see the helpful message, your code works as expected!
P.S. I have tested gradient checking for my solutions on all files. Therefore, you do not need to change the testing scripts at all in order to get a working solution (aka no more BCE scenarios)! Please come to office hours for help if you need it!
1