File size: 9,969 Bytes
2c8d22f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
""" Adafactor Optimizer

Lifted from https://github.com/pytorch/fairseq/blob/master/fairseq/optim/adafactor.py

Modified by Ross Wightman to fix some issues with factorization dims for non nn.Linear layers

Original header/copyright below.
"""
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import math
from typing import Optional, Tuple

import torch

from ._types import ParamsT


class Adafactor(torch.optim.Optimizer):
    """Implements Adafactor algorithm.

    This implementation is based on: `Adafactor: Adaptive Learning Rates with Sublinear Memory Cost`
    (see https://arxiv.org/abs/1804.04235)

    Note that this optimizer internally adjusts the learning rate depending on the
    *scale_parameter*, *relative_step* and *warmup_init* options.

    To use a manual (external) learning rate schedule you should set `scale_parameter=False` and
    `relative_step=False`.

    Ags:
        params: iterable of parameters to optimize or dicts defining parameter groups
        lr: external learning rate
        eps: regularization constants for square gradient and parameter scale respectively
        eps_scale: regularization constants for parameter scale respectively
        clip_threshold: threshold of root-mean-square of final gradient update
        decay_rate: coefficient used to compute running averages of square gradient
        beta1: coefficient used for computing running averages of gradient
        weight_decay: weight decay
        scale_parameter: if True, learning rate is scaled by root-mean-square of parameter
        warmup_init: time-dependent learning rate computation depends on whether warm-up initialization is being used
    """

    def __init__(
            self,
            params: ParamsT,
            lr: Optional[float] = None,
            eps: float = 1e-30,
            eps_scale: float = 1e-3,
            clip_threshold: float = 1.0,
            decay_rate: float = -0.8,
            betas: Optional[Tuple[float, float]] = None,
            weight_decay: float = 0.0,
            scale_parameter: bool = True,
            warmup_init: bool = False,
            min_dim_size_to_factor: int = 16,
            caution: bool = False,
    ):
        relative_step = not lr
        if warmup_init and not relative_step:
            raise ValueError('warmup_init requires relative_step=True')

        beta1 = None if betas is None else betas[0]   # make it compat with standard betas arg
        defaults = dict(
            lr=lr,
            eps=eps,
            eps_scale=eps_scale,
            clip_threshold=clip_threshold,
            decay_rate=decay_rate,
            beta1=beta1,
            weight_decay=weight_decay,
            scale_parameter=scale_parameter,
            relative_step=relative_step,
            warmup_init=warmup_init,
            min_dim_size_to_factor=min_dim_size_to_factor,
            caution=caution,
        )
        super(Adafactor, self).__init__(params, defaults)

    def __setstate__(self, state):
        super().__setstate__(state)
        for group in self.param_groups:
            group.setdefault('caution', False)
            group.setdefault('min_dim_size_to_factor', 16)

    @staticmethod
    def _get_lr(param_group, param_state):
        if param_group['relative_step']:
            min_step = 1e-6 * param_state['step'] if param_group['warmup_init'] else 1e-2
            lr_t = min(min_step, 1.0 / math.sqrt(param_state['step']))
            param_scale = 1.0
            if param_group['scale_parameter']:
                param_scale = max(param_group['eps_scale'], param_state['RMS'])
            param_group['lr'] = lr_t * param_scale
        return param_group['lr']

    @staticmethod
    def _get_options(param_group, param_shape, min_size_to_factor=16):
        use_first_moment = param_group['beta1'] is not None
        factored = None
        ndim = len(param_shape)
        # Use a simple heuristic to pick factorization row & col, note other PyTorch impl tend to
        # always use -2, -1 BUT this will not pick correct dims for convolutions. This is a simple
        # approach that should work in most cases, compare to the slightly more involved approach
        # in AdafactorBigVision that sorts dims by size, please report if wrong dims chosen.
        if ndim > 2 and param_shape[0] > min_size_to_factor and param_shape[1] > min_size_to_factor:
            # nD convs in torch are ND + 2 dim weights with leading in/out chs
            factored = 0, 1
        elif ndim >= 2 and param_shape[-2] > min_size_to_factor and param_shape[-1] > min_size_to_factor:
            # if the criteria above didn't match, test trailing dims for eligibility as per original impl
            factored = ndim - 2, ndim - 1

        return factored, use_first_moment

    @staticmethod
    def _rms(tensor):
        return tensor.norm(2) / (tensor.numel() ** 0.5)

    def _approx_sq_grad(self, exp_avg_sq_row, exp_avg_sq_col, dim_col, dim_row):
        # from our dim heuristic, always dim_col < dim_row, so col reduction dim for factored row = dim_col
        r_factor = (exp_avg_sq_row / exp_avg_sq_row.mean(dim=dim_col, keepdim=True)).rsqrt_().unsqueeze(dim_row)
        c_factor = exp_avg_sq_col.unsqueeze(dim_col).rsqrt()
        return torch.mul(r_factor, c_factor)

    @torch.no_grad()
    def step(self, closure=None):
        """Performs a single optimization step.
        Arguments:
            closure (callable, optional): A closure that reevaluates the model and returns the loss.
        """
        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            for p in group['params']:
                if p.grad is None:
                    continue
                grad = p.grad
                if grad.dtype in {torch.float16, torch.bfloat16}:
                    grad = grad.float()
                if grad.is_sparse:
                    raise RuntimeError('Adafactor does not support sparse gradients.')

                state = self.state[p]

                factored_dims, use_first_moment = self._get_options(
                    group,
                    grad.shape,
                    min_size_to_factor=group['min_dim_size_to_factor'],
                )
                # State Initialization
                if len(state) == 0:
                    state['step'] = 0

                    if use_first_moment:
                        # Exponential moving average of gradient values
                        state['exp_avg'] = torch.zeros_like(grad)
                    if factored_dims is not None:
                        dim_col, dim_row = factored_dims
                        def _remove_dim(shape, dim):
                            return shape[:dim] + shape[dim + 1:]
                        state['exp_avg_sq_row'] = torch.zeros(_remove_dim(grad.shape, dim_row)).to(grad)
                        state['exp_avg_sq_col'] = torch.zeros(_remove_dim(grad.shape, dim_col)).to(grad)
                    else:
                        state['exp_avg_sq'] = torch.zeros_like(grad)

                    state['RMS'] = 0
                else:
                    if use_first_moment:
                        state['exp_avg'] = state['exp_avg'].to(grad)
                    if factored_dims is not None:
                        state['exp_avg_sq_row'] = state['exp_avg_sq_row'].to(grad)
                        state['exp_avg_sq_col'] = state['exp_avg_sq_col'].to(grad)
                    else:
                        state['exp_avg_sq'] = state['exp_avg_sq'].to(grad)

                p_fp32 = p
                if p.dtype in {torch.float16, torch.bfloat16}:
                    p_fp32 = p_fp32.float()

                state['step'] += 1
                state['RMS'] = self._rms(p_fp32)
                lr_t = self._get_lr(group, state)

                beta2t = 1.0 - math.pow(state['step'], group['decay_rate'])
                update = grad ** 2 + group['eps']
                if factored_dims is not None:
                    dim_col, dim_row = factored_dims
                    exp_avg_sq_row = state['exp_avg_sq_row']
                    exp_avg_sq_col = state['exp_avg_sq_col']

                    exp_avg_sq_row.mul_(beta2t).add_(update.mean(dim=dim_row), alpha=1.0 - beta2t)
                    exp_avg_sq_col.mul_(beta2t).add_(update.mean(dim=dim_col), alpha=1.0 - beta2t)

                    # Approximation of exponential moving average of square of gradient
                    update = self._approx_sq_grad(exp_avg_sq_row, exp_avg_sq_col, dim_col, dim_row)
                    update.mul_(grad)
                else:
                    exp_avg_sq = state['exp_avg_sq']

                    exp_avg_sq.mul_(beta2t).add_(update, alpha=1.0 - beta2t)
                    update = exp_avg_sq.rsqrt().mul_(grad)

                update.div_((self._rms(update) / group['clip_threshold']).clamp_(min=1.0))
                update.mul_(lr_t)

                if use_first_moment:
                    exp_avg = state['exp_avg']
                    exp_avg.mul_(group['beta1']).add_(update, alpha=1 - group['beta1'])
                    if group['caution']:
                        # Apply caution as per 'Cautious Optimizers' - https://arxiv.org/abs/2411.16085
                        mask = (exp_avg * grad > 0).to(grad.dtype)
                        mask.div_(mask.mean().clamp_(min=1e-3))
                        update = exp_avg * mask
                    else:
                        update = exp_avg

                if group['weight_decay'] != 0:
                    p_fp32.add_(p_fp32, alpha=-group['weight_decay'] * lr_t)

                p_fp32.add_(-update)
                if p.dtype in {torch.float16, torch.bfloat16}:
                    p.copy_(p_fp32)

        return loss