summaryrefslogtreecommitdiffstats
path: root/training/optimization.py
blob: 7d8d55a40d26f09c0e7a774a7a41f964c20d3d5b (plain) (blame)
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
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
import math
from typing import NamedTuple, Literal, Callable
from functools import partial

import torch
from torch.optim.lr_scheduler import LambdaLR

from diffusers.optimization import get_scheduler as get_scheduler_, get_cosine_with_hard_restarts_schedule_with_warmup


class OneCyclePhase(NamedTuple):
    step_min: int
    step_max: int
    min: float
    max: float
    func: Callable[[float], float]


def warmup_linear(progress: float):
    return progress


def warmup_cos(exp: int, progress: float):
    lr = 0.5 * (1.0 + math.cos(math.pi * (1 + progress)))
    lr = lr ** (exp - (exp - 1) * progress)
    return lr


def anneal_linear(progress: float):
    return 1 - progress


def anneal_half_cos(exp: int, progress: float):
    lr = 1.0 + math.cos(math.pi * (0.5 + 0.5 * progress))
    lr = lr ** (exp - (exp - 1) * progress)
    return lr


def anneal_cos(exp: int, progress: float):
    lr = 0.5 * (1.0 + math.cos(math.pi * progress))
    lr = lr ** (exp - (exp - 1) * progress)
    return lr


def get_one_cycle_schedule(
    optimizer: torch.optim.Optimizer,
    num_training_steps: int,
    warmup: Literal["cos", "linear"] = "cos",
    annealing: Literal["cos", "half_cos", "linear"] = "cos",
    warmup_exp: int = 1,
    annealing_exp: int = 1,
    min_lr: float = 0.04,
    mid_point: float = 0.3,
    last_epoch: int = -1
):
    if warmup == "linear":
        warmup_func = warmup_linear
    else:
        warmup_func = partial(warmup_cos, warmup_exp)

    if annealing == "linear":
        anneal_func = anneal_linear
    elif annealing == "half_cos":
        anneal_func = partial(anneal_half_cos, annealing_exp)
    else:
        anneal_func = partial(anneal_cos, annealing_exp)

    thresh_up = int(num_training_steps * min(mid_point, 0.5))

    if annealing == "linear":
        thresh_down = thresh_up * 2
        phases = [
            OneCyclePhase(0, thresh_up, min_lr, 1, warmup_func),
            OneCyclePhase(thresh_up, thresh_down, min_lr, 1, anneal_func),
            OneCyclePhase(thresh_down, num_training_steps, 0, min_lr, anneal_func),
        ]
    else:
        phases = [
            OneCyclePhase(0, thresh_up, min_lr, 1, warmup_func),
            OneCyclePhase(thresh_up, num_training_steps, 0, 1, anneal_func),
        ]

    def lr_lambda(current_step: int):
        phase = [p for p in phases if current_step >= p.step_min][-1]
        return phase.min + phase.func((current_step - phase.step_min) / (phase.step_max - phase.step_min)) * (phase.max - phase.min)

    return LambdaLR(optimizer, lr_lambda, last_epoch)


def get_exponential_growing_schedule(optimizer, end_lr: float, num_training_steps: int, last_epoch: int = -1):
    def lr_lambda(base_lr: float, current_step: int):
        return (end_lr / base_lr) ** (current_step / num_training_steps)

    lr_lambdas = [partial(lr_lambda, group["lr"]) for group in optimizer.param_groups]

    return LambdaLR(optimizer, lr_lambdas, last_epoch)


def get_scheduler(
    id: str,
    optimizer: torch.optim.Optimizer,
    num_training_steps_per_epoch: int,
    gradient_accumulation_steps: int = 1,
    min_lr: float = 0.04,
    warmup_func: Literal["cos", "linear"] = "cos",
    annealing_func: Literal["cos", "half_cos", "linear"] = "cos",
    warmup_exp: int = 1,
    annealing_exp: int = 1,
    end_lr: float = 1e3,
    cycles: int = 1,
    train_epochs: int = 100,
    warmup_epochs: int = 10,
):
    num_training_steps_per_epoch = math.ceil(
        num_training_steps_per_epoch / gradient_accumulation_steps
    )  # * gradient_accumulation_steps
    num_training_steps = train_epochs * num_training_steps_per_epoch
    num_warmup_steps = warmup_epochs * num_training_steps_per_epoch

    if id == "one_cycle":
        lr_scheduler = get_one_cycle_schedule(
            optimizer=optimizer,
            num_training_steps=num_training_steps,
            warmup=warmup_func,
            annealing=annealing_func,
            warmup_exp=warmup_exp,
            annealing_exp=annealing_exp,
            min_lr=min_lr,
        )
    elif id == "exponential_growth":
        if cycles is None:
            cycles = math.ceil(math.sqrt(((num_training_steps - num_warmup_steps) / num_training_steps_per_epoch)))

        lr_scheduler = get_exponential_growing_schedule(
            optimizer=optimizer,
            end_lr=end_lr,
            num_training_steps=num_training_steps,
        )
    elif id == "cosine_with_restarts":
        if cycles is None:
            cycles = math.ceil(math.sqrt(((num_training_steps - num_warmup_steps) / num_training_steps_per_epoch)))

        lr_scheduler = get_cosine_with_hard_restarts_schedule_with_warmup(
            optimizer=optimizer,
            num_warmup_steps=num_warmup_steps,
            num_training_steps=num_training_steps,
            num_cycles=cycles,
        )
    else:
        lr_scheduler = get_scheduler_(
            id,
            optimizer=optimizer,
            num_warmup_steps=num_warmup_steps,
            num_training_steps=num_training_steps,
        )

    return lr_scheduler