Cosine Annealing with Warm Restarts

Implement a learning rate scheduler based on the SGDR (Stochastic Gradient Descent with Warm Restarts) technique from Loshchilov & Hutter, 2017. The scheduler uses cosine annealing to decay the learning rate from a maximum value to a minimum value within each cycle. At the end of each cycle, the learning rate "restarts" back to the maximum. The cycle length can optionally grow after each restart by a multiplicative factor. Write a function cosine annealing warm restarts(eta max, eta min, T 0, T mult, total epochs) that computes the learning rate for each epoch from 0 to total epochs 1. Parameters: eta max (float): Maximum (initial) learning rate at the start of each cycle eta min (float): Minimum learning rate at the end of each cycle T 0 (int): Number of epochs in the first cycle T mult (int): Multiplicative factor for cycle length after each restart. The second cycle has length T 0 T mult, the third has length T 0 T mult^2, and so on. When T mult is 1, all cycles have the same length T 0. total epochs (int): Total number of epochs to compute learning rates for The function should return a list of floats, where each float is the learning rate for the corresponding epoch, rounded to 4 decimal places. For each epoch, you need to determine which cycle it belongs to, compute how far into that cycle it is, and apply the cosine annealing formula accordingly.