| """ |
| Author: Mélanie Gaillochet |
| Date: 2020-10-27 |
| """ |
| from torch.optim import lr_scheduler |
|
|
| schedulers = { |
| 'StepLR': lr_scheduler.StepLR, |
| 'MultiStepLR': lr_scheduler.MultiStepLR, |
| 'ReduceOnPlateau': lr_scheduler.ReduceLROnPlateau, |
| 'CosineAnnealingLR': lr_scheduler.CosineAnnealingLR |
| |
| } |
|
|
| """ |
| Variables to be defined in config |
| StepLR: { |
| "step_size" (decreased LR every k epochs) |
| "gamma" (factor by which LR will be reduced) |
| }, |
| |
| MultiStepLR: { |
| "milestones" (list of epochs where the LR is decreased) |
| "gamma" (factor by which LR will be reduced) |
| }, |
| |
| ReduceOnPlateau: { |
| "mode" (ie: "max", "min") - LR reduced when quantity has reached max or min |
| "factor" (default 0.1) - Factor by which LR will be reduced |
| "patience": (default 10) - # epochs with no improvement after which LR reduced |
| "verbose": (ie: True, False) |
| "min_lr": (default 0) - Lower bound on LR |
| unchanged "threshold" (default 1e-4) - Threshold for measuring the new optimum, |
| } |
| """ |
|
|
