optimization.py

# coding=utf-8
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""PyTorch optimization for BERT model."""

import logging
import math

import torch
from torch.optim import Optimizer
from torch.optim.lr_scheduler import LambdaLR

logger = logging.getLogger(__name__)

class ConstantLRSchedule(LambdaLR):
    def __init__(self, optimizer, last_epoch=-1):
        super(ConstantLR, self).__init__(optimizer, lambda x: x, last_epoch=last_epoch)

class WarmupCosineSchedule(LambdaLR):
    """
    Linearly increases learning rate from 0 to 1 over `warmup` training steps.
    Decreases learning rate from 1. to 0. over remaining `t_total - warmup` steps following a cosine curve.
    If `cycles` (default=0.5) is different from default, learning rate follows cosine function after warmup.
    :param warmup:      see LRSchedule
    :param t_total:     see LRSchedule
    :param cycles:      number of cycles. Default: 0.5, corresponding to cosine decay from 1. at progress==warmup and 0 at progress==1.
    :param kw:
    """
    warn_t_total = True
    def __init__(self, optimizer, warmup_steps, t_total, cycles=.5, last_epoch=-1):

        def lr_lambda(step):
            if step < warmup_steps:
                return step / max(1, warmup_steps)
            else:
                progress = (step - warmup_steps) / max(1, t_total - warmup_steps)   # progress after warmup
                return 0.5 * (1. + math.cos(math.pi * cycles * 2 * progress))

        super(WarmupCosineSchedule, self).__init__(optimizer, lr_lambda, last_epoch=last_epoch)

class WarmupCosineWithHardRestartsSchedule(LambdaLR):
    """
    Linearly increases learning rate from 0 to 1 over `warmup` fraction of training steps.
    If `cycles` (default=1.) is different from default, learning rate follows `cycles` times a cosine decaying
    learning rate (with hard restarts).
    """
    def __init__(self, optimizer, warmup_steps, t_total, cycles=1., last_epoch=-1):

        def lr_lambda(step):
            if step < warmup_steps:
                return step / max(1, warmup_steps)
            else:
                progress = (step - warmup_steps) / max(1, t_total - warmup_steps)   # progress after warmup
                ret = 0.5 * (1. + math.cos(math.pi * ((cycles * progress) % 1)))
                return ret

        super(WarmupCosineWithHardRestartsSchedule, self).__init__(optimizer, lr_lambda, last_epoch=last_epoch)


class WarmupConstantSchedule(LambdaLR):
    """
    Linearly increases learning rate from 0 to 1 over `warmup` fraction of training steps.
    Keeps learning rate equal to 1. after warmup.
    """
    def __init__(self, optimizer, warmup_steps, last_epoch=-1):

        def lr_lambda(step):
            if step < warmup_steps:
                return step / warmup_steps
            return 1.

        super(WarmupConstantSchedule, self).__init__(optimizer, lr_lambda, last_epoch=last_epoch)


class WarmupLinearSchedule(LambdaLR):
    """
    Linearly increases learning rate from 0 to 1 over `warmup` fraction of training steps.
    Linearly decreases learning rate from 1. to 0. over remaining `1 - warmup` steps.
    """
    def __init__(self, optimizer, warmup_steps, t_total, last_epoch=-1):

        def lr_lambda(step):
            if step < warmup_steps:
                return step / max(1, warmup_steps)
            return (t_total - step) / max(1, t_total - warmup_steps)

        super(WarmupLinearSchedule, self).__init__(optimizer, lr_lambda, last_epoch=last_epoch)


class AdamW(Optimizer):
    """ Implements Adam algorithm with weight decay fix.

    Parameters:
        lr: learning rate
        warmup: portion of t_total for the warmup, -1  means no warmup. Default: -1
        t_total: total number of training steps for the learning
            rate schedule, -1  means constant learning rate of 1. (no warmup regardless of warmup setting). Default: -1
        schedule: schedule to use for the warmup (see above).
            Can be `'warmup_linear'`, `'warmup_constant'`, `'warmup_cosine'`, `'none'`, `None` or a `_LRSchedule` object (see below).
            If `None` or `'none'`, learning rate is always kept constant.
            Default : `'warmup_linear'`
        b1: Adams b1. Default: 0.9
        b2: Adams b2. Default: 0.999
        e: Adams epsilon. Default: 1e-6
        weight_decay: Weight decay. Default: 0.01
        max_grad_norm: Maximum norm for the gradients (-1 means no clipping). Default: 1.0
        correct_bias: can be set to False to avoid correcting bias in Adam (e.g. like in Bert repository)
    """
    def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-6, weight_decay=0.01, correct_bias=True):
        if lr < 0.0:
            raise ValueError("Invalid learning rate: {} - should be >= 0.0".format(lr))
        if not 0.0 <= betas[0] < 1.0:
            raise ValueError("Invalid beta parameter: {} - should be in [0.0, 1.0[".format(betas[0]))
        if not 0.0 <= betas[1]  < 1.0:
            raise ValueError("Invalid beta parameter: {} - should be in [0.0, 1.0[".format(betas[1] ))
        if not 0.0 <= eps:
            raise ValueError("Invalid epsilon value: {} - should be >= 0.0".format(e))
        defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay,
                        correct_bias=correct_bias)
        super(BertAdam, self).__init__(params, defaults)

    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')

                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)

                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
                beta1, beta2 = group['betas']

                state['step'] += 1

                # Decay the first and second moment running average coefficient
                # In-place operations to update the averages at the same time
                exp_avg.mul_(beta1).add_(1 - beta1, grad)
                exp_avg_sq.mul_(beta2).addcmul_(1 - beta2, grad, grad)
                denom = exp_avg_sq.sqrt().add_(group['eps'])

                step_size = group['lr']
                if group['correct_bias']:  # No bias correction for Bert
                    bias_correction1 = 1 - beta1 ** state['step']
                    bias_correction2 = 1 - beta2 ** state['step']
                    step_size = step_size * math.sqrt(bias_correction2) / bias_correction1

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

                # Just adding the square of the weights to the loss function is *not*
                # the correct way of using L2 regularization/weight decay with Adam,
                # since that will interact with the m and v parameters in strange ways.
                #
                # Instead we want to decay the weights in a manner that doesn't interact
                # with the m/v parameters. This is equivalent to adding the square
                # of the weights to the loss with plain (non-momentum) SGD.
                # Add weight decay at the end (fixed version)
                if group['weight_decay'] > 0:
                    p.data.add_(-group['lr'] * group['weight_decay'], p.data)

        return loss