File size: 1,747 Bytes
f8f62f3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
50
51
52
53
54
import numpy as np


def assign_learning_rate(optimizer, new_lr):
    for param_group in optimizer.param_groups:
        param_group["lr"] = new_lr


def _warmup_lr(base_lr, warmup_length, step):
    return base_lr * (step + 1) / warmup_length


def const_lr(optimizer, base_lr, warmup_length, steps):
    def _lr_adjuster(step):
        if step < warmup_length:
            lr = _warmup_lr(base_lr, warmup_length, step)
        else:
            lr = base_lr
        assign_learning_rate(optimizer, lr)
        return lr
    return _lr_adjuster


def const_lr_cooldown(optimizer, base_lr, warmup_length, steps, cooldown_steps, cooldown_power=1.0, cooldown_end_lr=0.):
    def _lr_adjuster(step):
        start_cooldown_step = steps - cooldown_steps
        if step < warmup_length:
            lr = _warmup_lr(base_lr, warmup_length, step)
        else:
            if step < start_cooldown_step:
                lr = base_lr
            else:
                e = step - start_cooldown_step
                es = steps - start_cooldown_step
                # linear decay if power == 1; polynomial decay otherwise;
                decay = (1 - (e/es)) ** cooldown_power
                lr = decay * (base_lr - cooldown_end_lr) + cooldown_end_lr
        assign_learning_rate(optimizer, lr)
        return lr
    return _lr_adjuster


def cosine_lr(optimizer, base_lr, warmup_length, steps):
    def _lr_adjuster(step):
        if step < warmup_length:
            lr = _warmup_lr(base_lr, warmup_length, step)
        else:
            e = step - warmup_length
            es = steps - warmup_length
            lr = 0.5 * (1 + np.cos(np.pi * e / es)) * base_lr
        assign_learning_rate(optimizer, lr)
        return lr
    return _lr_adjuster