In this assignment, it would be your task to implement your own autograd library, similar to the one shown in the lab. The difference is that while in the lab, we showed autograd for scalar values, your solution shall also work with vectors and tensors.
All necessary files are provided in the assignment here.
The task is to complete the engine.py file. Here, you will find ellipses (…), where you should complete the functions to compute forward and backward passes correctly. There are also starting and ending with recognizable comments. Inside them, you should provide your own code. These are used for longer sections of code. In this homework, they are inside higher-level functions of the autograd.
You will be using the numpy library, and its documentation may be found here . Do NOT use any libraries with autograd, such as PyTorch or Tensorflow.
Most of the functions should be elementary. However, if you are unsure how to proceed with more complex ones, consult the lab materials before contacting tutors.
Please be aware that the variable other in function __pow__ is either int or float and, therefore, cannot be put in the parent set of output. Backpropagation is available only for objects in the class Tensor.
Another tip is that the reshape_gradient function is only needed in __add__ and __mul__ functions. Only in these two functions do we have the possibility of broadcasting. The reshape_gradient function aims to solve the problem of calculating the gradient for broadcasted values.
You may use the provided test.py file to test your solution. Similar but slightly more complex testing is used in BRUTE.
The whole directory with all files provided in the assignment must be zipped and uploaded to BRUTE. You may use the provided submit.sh file. The deadline for the submission is 03.11.2024 23:59:59.
The maximum amount of points you may get is 10. The whole assignment will be auto-evaluated in BRUTE. The tutors may later re-evaluate any submission.
You will receive minus one point for every 24 hours after the deadline. However, no more than 9 points will be deducted for late submission.
Good luck, do not forget to play with the task a bit, and in the case of any questions or concerns please contact vlkjan6@fel.cvut.cz.