File size: 9,184 Bytes
6a62ffb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
231
232
233
234
235
236
237
238
239
240
# 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 logging
import math
from collections.abc import Collection
from dataclasses import dataclass, field
from typing import Any, List

import torch
import torch.distributed as dist
import torch.optim
from fairseq.dataclass import FairseqDataclass
from fairseq.optim import FairseqOptimizer, register_optimizer
from fairseq.optim.fused_adam import get_fused_adam_class
from omegaconf import II, OmegaConf


logger = logging.getLogger(__name__)


@dataclass
class FairseqAdamConfig(FairseqDataclass):
    adam_betas: Any = field(
        default=(0.9, 0.999), metadata={"help": "betas for Adam optimizer"}
    )
    adam_eps: float = field(
        default=1e-8, metadata={"help": "epsilon for Adam optimizer"}
    )
    weight_decay: float = field(default=0.0, metadata={"help": "weight decay"})
    use_old_adam: bool = field(
        default=False, metadata={"help": "Use fairseq.optim.adam.Adam"}
    )
    fp16_adam_stats: bool = field(
        default=False, metadata={"help": "use FP16 stats (with automatic scaling)"}
    )
    # TODO common vars below in parent
    tpu: bool = II("common.tpu")
    lr: List[float] = II("optimization.lr")


@register_optimizer("adam", dataclass=FairseqAdamConfig)
class FairseqAdam(FairseqOptimizer):
    """Adam optimizer for fairseq.

    Important note: this optimizer corresponds to the "AdamW" variant of
    Adam in its weight decay behavior. As such, it is most closely
    analogous to torch.optim.AdamW from PyTorch.
    """

    def __init__(self, cfg: FairseqAdamConfig, params):
        super().__init__(cfg)
        fused_adam_cls = get_fused_adam_class()
        use_fused_adam = (
            not getattr(cfg, "use_old_adam", False)
            and fused_adam_cls is not None
            and torch.cuda.is_available()
        )
        if getattr(cfg, "tpu", False):
            if self.cfg.fp16_adam_stats:
                raise NotImplementedError("--fp16-adam-stats is only supported on GPU")
            # on TPUs we use the Adam defined here, since it
            # automatically casts gradients to FP32
            self._optimizer = Adam(params, **self.optimizer_config)
        elif use_fused_adam:
            logger.info("using FusedAdam")
            self._optimizer = fused_adam_cls(
                params, use_fp16_stats=self.cfg.fp16_adam_stats, **self.optimizer_config
            )
        else:
            if self.cfg.fp16_adam_stats:
                raise NotImplementedError(
                    "--fp16-adam-stats is only supported with FusedAdamV1"
                )
            self._optimizer = Adam(params, **self.optimizer_config)

    @property
    def optimizer_config(self):
        """
        Return a kwarg dictionary that will be used to override optimizer
        args stored in checkpoints. This allows us to load a checkpoint and
        resume training using a different set of optimizer args, e.g., with a
        different learning rate.
        """
        return {
            "lr": self.cfg.lr[0]
            if isinstance(self.cfg.lr, Collection)
            else self.cfg.lr,
            "betas": eval(self.cfg.adam_betas)
            if isinstance(self.cfg.adam_betas, str)
            else OmegaConf.to_container(self.cfg.adam_betas),
            "eps": self.cfg.adam_eps,
            "weight_decay": self.cfg.weight_decay,
        }

    def average_params(self):
        """Reduce Params is only used during BMUF distributed training."""
        state_dict = self.optimizer.state_dict()
        total_gpus = float(dist.get_world_size())

        for _, value in state_dict["state"].items():
            value["exp_avg"] /= total_gpus
            value["exp_avg_sq"] /= total_gpus
            dist.all_reduce(value["exp_avg"], op=dist.ReduceOp.SUM)
            dist.all_reduce(value["exp_avg_sq"], op=dist.ReduceOp.SUM)


class Adam(torch.optim.Optimizer):
    r"""Implements Adam algorithm.

    This implementation is modified from torch.optim.Adam based on:
    `Fixed Weight Decay Regularization in Adam`
    (see https://arxiv.org/abs/1711.05101)

    It has been proposed in `Adam: A Method for Stochastic Optimization`_.

    Args:
        params (iterable): iterable of parameters to optimize or dicts defining
            parameter groups
        lr (float, optional): learning rate (default: 1e-3)
        betas (Tuple[float, float], optional): coefficients used for computing
            running averages of gradient and its square (default: (0.9, 0.999))
        eps (float, optional): term added to the denominator to improve
            numerical stability (default: 1e-8)
        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
        amsgrad (boolean, optional): whether to use the AMSGrad variant of this
            algorithm from the paper `On the Convergence of Adam and Beyond`_

    .. _Adam\: A Method for Stochastic Optimization:
        https://arxiv.org/abs/1412.6980
    .. _On the Convergence of Adam and Beyond:
        https://openreview.net/forum?id=ryQu7f-RZ
    """

    def __init__(
        self,
        params,
        lr=1e-3,
        betas=(0.9, 0.999),
        eps=1e-8,
        weight_decay=0,
        amsgrad=False,
    ):
        defaults = dict(
            lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, amsgrad=amsgrad
        )
        super(Adam, self).__init__(params, defaults)

    @property
    def supports_memory_efficient_fp16(self):
        return True

    @property
    def supports_flat_params(self):
        return True

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

        Args:
            closure (callable, optional): A closure that reevaluates the model
                and returns the loss.
        """
        loss = None
        if closure is not None:
            loss = closure()

        for group in self.param_groups:
            for p in group["params"]:
                if p.grad is None:
                    continue
                grad = p.grad.data
                if grad.dtype in {torch.float16, torch.bfloat16}:
                    grad = grad.float()
                if grad.is_sparse:
                    raise RuntimeError(
                        "Adam does not support sparse gradients, please consider SparseAdam instead"
                    )
                amsgrad = group.get("amsgrad", False)

                p_data_fp32 = p.data
                if p.data.dtype in {torch.float16, torch.bfloat16}:
                    p_data_fp32 = p_data_fp32.float()

                state = self.state[p]

                # State initialization
                if len(state) == 0:
                    state["step"] = 0
                    # Exponential moving average of gradient values
                    state["exp_avg"] = torch.zeros_like(p_data_fp32)
                    # Exponential moving average of squared gradient values
                    state["exp_avg_sq"] = torch.zeros_like(p_data_fp32)
                    if amsgrad:
                        # Maintains max of all exp. moving avg. of sq. grad. values
                        state["max_exp_avg_sq"] = torch.zeros_like(p_data_fp32)
                else:
                    state["exp_avg"] = state["exp_avg"].to(p_data_fp32)
                    state["exp_avg_sq"] = state["exp_avg_sq"].to(p_data_fp32)
                    if amsgrad:
                        state["max_exp_avg_sq"] = state["max_exp_avg_sq"].to(
                            p_data_fp32
                        )

                exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
                if amsgrad:
                    max_exp_avg_sq = state["max_exp_avg_sq"]
                beta1, beta2 = group["betas"]

                state["step"] += 1

                # Decay the first and second moment running average coefficient
                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
                if amsgrad:
                    # Maintains the maximum of all 2nd moment running avg. till now
                    torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)
                    # Use the max. for normalizing running avg. of gradient
                    denom = max_exp_avg_sq.sqrt().add_(group["eps"])
                else:
                    denom = exp_avg_sq.sqrt().add_(group["eps"])

                bias_correction1 = 1 - beta1 ** state["step"]
                bias_correction2 = 1 - beta2 ** state["step"]
                step_size = group["lr"] * math.sqrt(bias_correction2) / bias_correction1

                if group["weight_decay"] != 0:
                    p_data_fp32.add_(
                        p_data_fp32, alpha=-group["weight_decay"] * group["lr"]
                    )

                p_data_fp32.addcdiv_(exp_avg, denom, value=-step_size)

                if p.data.dtype in {torch.float16, torch.bfloat16}:
                    p.data.copy_(p_data_fp32)

        return loss