File size: 4,920 Bytes
3cd1076
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python
# -*- coding: utf-8 -*-

import torch
from torch.optim.optimizer import Optimizer

class LARS(Optimizer):
    r"""Extends SGD in PyTorch with LARS scaling from the paper
    `Large batch training of Convolutional Networks`__.
    .. note::
        The application of momentum in the SGD part is modified according to
        the PyTorch standards. LARS scaling fits into the equation in the
        following fashion.

        .. math::
            \begin{aligned}
                g_{t+1} & = \text{lars_lr} * (\beta * p_{t} + g_{t+1}), \\
                v_{t+1} & = \\mu * v_{t} + g_{t+1}, \\
                p_{t+1} & = p_{t} - \text{lr} * v_{t+1},
            \\end{aligned}

        where :math:`p`, :math:`g`, :math:`v`, :math:`\\mu` and :math:`\beta`
        denote the parameters, gradient, velocity, momentum, and weight decay
        respectively.  The :math:`lars_lr` is defined by Eq. 6 in the paper.
        The Nesterov version is analogously modified.

    .. warning::
        Parameters with weight decay set to 0 will automatically be excluded
        from layer-wise LR scaling. This is to ensure consistency with papers
        like SimCLR and BYOL.


    __ https://arxiv.org/pdf/1708.03888.pdf

    Note:
        Reference code: https://github.com/PyTorchLightning/lightning-bolts/
    """

    def __init__(
        self,
        params,
        lr: float = 1e-2,
        momentum: float = 0.0,
        dampening: float = 0.0,
        weight_decay: float = 0.0,
        nesterov: bool = False,
        trust_coefficient: float = 0.01,
        eps: float = 1e-8,
    ):
        if lr <= 0.0:
            raise ValueError("Invalid learning rate: {}".format(lr))
        if eps < 0.0:
            raise ValueError("Invalid epsilon value: {}".format(eps))
        if momentum < 0.0:
            raise ValueError("Invalid momentum value: {}".format(momentum))
        if dampening < 0.0:
            raise ValueError("Invalid dampening value: {}".format(dampening))
        if weight_decay < 0.0:
            raise ValueError(
                "Invalid weight_decay value: {}".format(weight_decay)
            )
        if trust_coefficient < 0.0:
            raise ValueError(
                "Invalid trust_coefficient value: {}".format(trust_coefficient)
            )

        defaults = dict(
            lr=lr,
            momentum=momentum,
            dampening=dampening,
            weight_decay=weight_decay,
            nesterov=nesterov,
            trust_coefficient=trust_coefficient,
            eps=eps,
        )
        if nesterov and (momentum <= 0 or dampening != 0):
            raise ValueError(
                "Nesterov momentum requires a momentum and zero dampening"
            )

        super().__init__(params, defaults)

    def __setstate__(self, state) -> None:
        super().__setstate__(state)

        for group in self.param_groups:
            group.setdefault("nesterov", False)

    @torch.no_grad()
    def step(self, closure = None):
        r"""Performs a single optimization step.

        Arguments:
            closure: A closure that reevaluates the model and returns the loss.
        """
        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        # exclude scaling for params with 0 weight decay
        for group in self.param_groups:
            weight_decay = group["weight_decay"]
            momentum = group["momentum"]
            dampening = group["dampening"]
            nesterov = group["nesterov"]

            for p in group["params"]:
                if p.grad is None:
                    continue

                d_p = p.grad
                p_norm = torch.norm(p.data)
                g_norm = torch.norm(p.grad.data)

                # lars scaling + weight decay part
                if weight_decay != 0:
                    if p_norm != 0 and g_norm != 0:
                        lars_lr = p_norm / (
                            g_norm + p_norm * weight_decay + group["eps"]
                        )
                        lars_lr *= group["trust_coefficient"]

                        d_p = d_p.add(p, alpha=weight_decay)
                        d_p *= lars_lr

                if momentum != 0:
                    param_state = self.state[p]
                    if "momentum_buffer" not in param_state:
                        buf = param_state["momentum_buffer"] = torch.clone(
                            d_p
                        ).detach()
                    else:
                        buf = param_state["momentum_buffer"]
                        buf.mul_(momentum).add_(d_p, alpha=1 - dampening)
                    if nesterov:
                        d_p = d_p.add(buf, alpha=momentum)
                    else:
                        d_p = buf

                p.add_(d_p, alpha=-group["lr"])

        return loss