| import numpy as np | |
| import matplotlib.pyplot as plt | |
| import numpy as np | |
| import matplotlib.pyplot as plt | |
| def triangle_lr_plot(lr_min, lr_max, step_size, iterations): | |
| lr_list = [] | |
| it_list = [j for j in range(iterations + 1)] | |
| for i in range(iterations + 1): | |
| half_cycle_count = np.floor(i/step_size) | |
| x = i - half_cycle_count * step_size | |
| if half_cycle_count % 2 == 0: | |
| lr = lr_min + x*(lr_max-lr_min)/step_size | |
| else: | |
| lr = lr_max - x*(lr_max-lr_min)/step_size | |
| lr_list.append(lr) | |
| fig, axs = plt.subplots(figsize=(5,5)) | |
| axs.plot(it_list, lr_list) | |
| plt.title("CLR - 'triangular' Policy") | |
| plt.xlabel("Iteration") | |
| plt.ylabel("Learning Rate") | |
| plt.show() | |
| def custom_one_cycle_lr(no_of_images, batch_size, base_lr, max_lr, final_lr, epoch_stage1=5, epoch_stage2=18, total_epochs=24): | |
| lr_schedule = lambda t: np.interp([t], [0, epoch_stage1, epoch_stage2, total_epochs], [base_lr, max_lr, base_lr, final_lr])[0] | |
| lr_lambda = lambda it: lr_schedule(it * batch_size/no_of_images) | |
| return lr_lambda | |
| def max_lr_finder_schedule(no_of_images, batch_size, base_lr, max_lr, total_epochs=5): | |
| lr_finder_schedule = lambda t: np.interp([t], [0, total_epochs], [base_lr, max_lr])[0] | |
| lr_finder_lambda = lambda it: lr_finder_schedule(it * batch_size/no_of_images) | |
| return lr_finder_lambda |