1. Part 1, My Perceptron: Is the input a scalar or list of floats?
Note: If your code is passing the autotests for Part 1 you do not need to make any changes.
The comment in the provided hw1.py file describing the functionality for my_perceptron() is incorrect. The functionality described in the assignment spec is what should be followed. The input, x, is a scalar that indicates the length of the placeholder to be created.
2. Memory Error when running train.py
This will occur if you try to train your model either ssh'd into a cse machine. You do not have to run train.py in the cse environment, you only need to verify that the autotests are passing. You should use train.py to verify locally that your models are reaching the expected accurracies.
The memory problem seems to occur on wagner and weill but not on williams. You can first log in to another cse machine and then type "ssh williams". Please try not to overload williams (i.e. use it only for final testing, or if you do not have your own laptop and are unable to physically go to a lab machine on campus).
3. Should my two-layer accuracy be higher than one-layer?
Yes, a correctly implemented two-layer network should acheive a higher accuracy. A common reason for this not to occur is an incorrect initialization scheme.