File size: 1,672 Bytes
4efbc62
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
# EfficientViT: Multi-Scale Linear Attention for High-Resolution Dense Prediction
# Han Cai, Junyan Li, Muyan Hu, Chuang Gan, Song Han
# International Conference on Computer Vision (ICCV), 2023

import math

import torch

from efficientvit.models.utils.list import val2list

__all__ = ["CosineLRwithWarmup"]


class CosineLRwithWarmup(torch.optim.lr_scheduler._LRScheduler):
    def __init__(
        self,
        optimizer: torch.optim.Optimizer,
        warmup_steps: int,
        warmup_lr: float,
        decay_steps: int or list[int],
        last_epoch: int = -1,
    ) -> None:
        self.warmup_steps = warmup_steps
        self.warmup_lr = warmup_lr
        self.decay_steps = val2list(decay_steps)
        super().__init__(optimizer, last_epoch)

    def get_lr(self) -> list[float]:
        if self.last_epoch < self.warmup_steps:
            return [
                (base_lr - self.warmup_lr) * (self.last_epoch + 1) / self.warmup_steps
                + self.warmup_lr
                for base_lr in self.base_lrs
            ]
        else:
            current_steps = self.last_epoch - self.warmup_steps
            decay_steps = [0] + self.decay_steps
            idx = len(decay_steps) - 2
            for i, decay_step in enumerate(decay_steps[:-1]):
                if decay_step <= current_steps < decay_steps[i + 1]:
                    idx = i
                    break
            current_steps -= decay_steps[idx]
            decay_step = decay_steps[idx + 1] - decay_steps[idx]
            return [
                0.5 * base_lr * (1 + math.cos(math.pi * current_steps / decay_step))
                for base_lr in self.base_lrs
            ]