File size: 9,421 Bytes
d758c99
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import attr
import math
import torch
import transformers

from seq2struct.utils import registry

registry.register('optimizer', 'adadelta')(torch.optim.Adadelta)
registry.register('optimizer', 'adam')(torch.optim.Adam)
registry.register('optimizer', 'sgd')(torch.optim.SGD)


@registry.register('lr_scheduler', 'warmup_polynomial')
@attr.s
class WarmupPolynomialLRScheduler:
    param_groups = attr.ib()
    num_warmup_steps = attr.ib()
    start_lr = attr.ib()
    end_lr = attr.ib()
    decay_steps = attr.ib()
    power = attr.ib()

    def update_lr(self, current_step):
        if current_step < self.num_warmup_steps:
            warmup_frac_done = current_step / self.num_warmup_steps
            new_lr = self.start_lr * warmup_frac_done
        else:
            new_lr = (
                (self.start_lr - self.end_lr) * (1 - (current_step - self.num_warmup_steps) / self.decay_steps) ** self.power
                + self.end_lr)

        for param_group in self.param_groups:
            param_group['lr'] = new_lr


@registry.register('lr_scheduler', 'warmup_polynomial_group')
@attr.s
class WarmupPolynomialLRSchedulerGroup(WarmupPolynomialLRScheduler):
    start_lrs = attr.ib()
    """
    Each param group has it's own start lr
    start lr is in the same order as param groups,
    """

    def update_lr(self, current_step):
        for start_lr, param_group in zip(self.start_lrs, self.param_groups):
            if current_step < self.num_warmup_steps:
                warmup_frac_done = current_step / self.num_warmup_steps
                new_lr = start_lr * warmup_frac_done
            else:
                new_lr = (
                    (start_lr - self.end_lr) * (1 - (current_step - self.num_warmup_steps) / self.decay_steps) ** self.power
                    + self.end_lr)

            param_group['lr'] = new_lr


@registry.register('lr_scheduler', 'warmup_cosine')
@attr.s
class WarmupCosineLRScheduler:
    param_groups = attr.ib()
    num_warmup_steps = attr.ib()
    start_lr = attr.ib()
    end_lr = attr.ib()
    decay_steps = attr.ib()

    def update_lr(self, current_step):
        if current_step < self.num_warmup_steps:
            warmup_frac_done = current_step / self.num_warmup_steps
            new_lr = self.start_lr * warmup_frac_done
        else:
            new_lr = (
                (self.start_lr - self.end_lr) * 0.5 * (1 + math.cos(math.pi * (current_step - self.num_warmup_steps) / self.decay_steps))
                + self.end_lr)

        for param_group in self.param_groups:
            param_group['lr'] = new_lr


@registry.register('lr_scheduler', 'noop')
class NoOpLRScheduler:
    def __init__(self, optimizer):
        pass

    def update_lr(self, current_step):
        pass

@registry.register('optimizer', 'bertAdamw')
class BertAdamW(transformers.AdamW):
    """
    Given a model and its bert module, create parameter groups with different lr
    """
    def __init__(self, non_bert_params, bert_params, lr=1e-3, bert_lr=2e-5, **kwargs):
        self.bert_param_group = {"params" : bert_params , "lr": bert_lr, "weight_decay": 0}
        self.non_bert_param_group = {"params" : non_bert_params} 

        params = [self.non_bert_param_group, self.bert_param_group]
        if "name" in kwargs: del kwargs["name"] #TODO: fix this
        super(BertAdamW, self).__init__(params, lr=lr, **kwargs)

@registry.register('lr_scheduler', 'bert_warmup_polynomial_group')
@attr.s
class BertWarmupPolynomialLRSchedulerGroup(WarmupPolynomialLRScheduler):
    """
    Set the lr of bert to be zero when the other param group is warming-up
    """
    start_lrs = attr.ib()

    # Bert parameters are in the second group by default
    def update_lr(self, current_step):
        for i, (start_lr, param_group) in enumerate(zip(self.start_lrs, self.param_groups)):
            if current_step < self.num_warmup_steps:
                if i == 0:
                    warmup_frac_done = current_step / self.num_warmup_steps
                    new_lr = start_lr * warmup_frac_done
                else: # fix bert during warm-up
                    assert i == 1
                    new_lr = 0
            else:
                new_lr = (
                    (start_lr - self.end_lr) * (1 - (current_step - self.num_warmup_steps) / self.decay_steps) ** self.power
                    + self.end_lr)

            param_group['lr'] = new_lr


@registry.register('optimizer', 'adamw')
class AdamW(torch.optim.Optimizer):
    """Implements Adam algorithm.
    It has been proposed in `Adam: A Method for Stochastic Optimization`_.
    Arguments:
        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

    **Modified to implement AdamW, see https://arxiv.org/pdf/1711.05101v3.pdf**
    """

    def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-8,
                 weight_decay=0, amsgrad=False):
        if not 0.0 <= lr:
            raise ValueError("Invalid learning rate: {}".format(lr))
        if not 0.0 <= eps:
            raise ValueError("Invalid epsilon value: {}".format(eps))
        if not 0.0 <= betas[0] < 1.0:
            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
        if not 0.0 <= betas[1] < 1.0:
            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
        defaults = dict(lr=lr, betas=betas, eps=eps,
                        weight_decay=weight_decay, amsgrad=amsgrad)
        super(AdamW, self).__init__(params, defaults)

    def __setstate__(self, state):
        super(AdamW, self).__setstate__(state)
        for group in self.param_groups:
            group.setdefault('amsgrad', False)

    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:
            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.is_sparse:
                    raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead')
                amsgrad = group['amsgrad']

                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)
                    # Exponential moving average of squared gradient values
                    state['exp_avg_sq'] = torch.zeros_like(p.data)
                    if amsgrad:
                        # Maintains max of all exp. moving avg. of sq. grad. values
                        state['max_exp_avg_sq'] = torch.zeros_like(p.data)

                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

                # MODIFIED HERE
                #if group['weight_decay'] != 0:
                #    grad = grad.add(group['weight_decay'], p.data)

                # Decay the first and second moment running average coefficient
                exp_avg.mul_(beta1).add_(1 - beta1, grad)
                exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
                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

                # MODIFIED HERE
                # Note that weight_decay is ultimately multiplied with the learning rate.
                update = exp_avg / denom
                if group['weight_decay'] != 0:
                    update += group['weight_decay'] * p.data
                p.data.add_(-step_size * update)

                #p.data.addcdiv_(-step_size, exp_avg, denom)

        return loss