bert_optim.py

# coding=utf-8
# Copyright (c) 2019 NVIDIA CORPORATION. All rights reserved.
# Copyright 2018 The Google AI Language Team Authors and The HugginFace 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 math
import torch
from torch.optim import Optimizer
from torch.optim.optimizer import required
from torch.nn.utils import clip_grad_norm_
from torch.optim.lr_scheduler import _LRScheduler


class LRScheduler(_LRScheduler):
    def __init__(self, optimizer, last_epoch=-1):
        # Check if using mixed precision training
        self.mixed_training = False
        base_optimizer = optimizer

        # Check that optimizer param is valid
        if not isinstance(optimizer, Optimizer):
            raise TypeError('{} is not an Optimizer'.format(
                type(optimizer).__name__))

        super(LRScheduler, self).__init__(base_optimizer, last_epoch)

    def step(self, epoch=None):
        # Set the current training step
        # ('epoch' is used to be consistent with _LRScheduler)
        if self.mixed_training:
            # The assumption is that the step will be constant
            state_dict = self.optimizer.state[self.optimizer.param_groups[0]['params'][0]]
            if 'step' in state_dict:
                self.last_epoch = state_dict['step'] + 1
            else:
                self.last_epoch = 1
        else:
            self.last_epoch = epoch if epoch is not None else self.last_epoch + 1

        for param_group, lr in zip(self.optimizer.param_groups, self.get_lr()):
            param_group['lr'] = lr


class PolyWarmUpScheduler(LRScheduler):
    """
    Applies a warm up period to the learning rate.
    """

    def __init__(self, optimizer, warmup, total_steps, degree=0.5, last_epoch=-1, base_lr=1., device='cpu'):
        self.warmup = torch.tensor(warmup, device=device)
        self.total_steps = torch.tensor(total_steps, device=device)
        self.degree = torch.tensor(degree, device=device)
        device_last_epoch = torch.tensor(last_epoch, device=device)
        self.base_lr = torch.tensor(base_lr, device=device)
        self.device = device
        super(PolyWarmUpScheduler, self).__init__(optimizer, device_last_epoch)

    def step(self, epoch=None):
        param_group = self.optimizer.param_groups[0]
        if 'step' in param_group:
            self.last_epoch = param_group['step'] + 1
        else:
            self.last_epoch = torch.tensor(1., device=self.device)

        for param_group, lr in zip(self.optimizer.param_groups, self.get_lr()):
            param_group['lr'] = lr

    def get_lr(self):
        progress = self.last_epoch / self.total_steps
        lr_tensor = torch.where(progress < self.warmup, self.base_lr * progress / self.warmup, self.base_lr * ((1.0 - progress) ** self.degree))
        return [lr_tensor for _ in range(len(self.optimizer.param_groups))]


def warmup_cosine(x, warmup=0.002):
    if x < warmup:
        return x/warmup
    return 0.5 * (1.0 + torch.cos(math.pi * x))


def warmup_constant(x, warmup=0.002):
    if x < warmup:
        return x/warmup
    return 1.0


def warmup_linear(x, warmup=0.002):
    if x < warmup:
        return x/warmup
    return max((x - 1.)/(warmup - 1.), 0.)


def warmup_poly(x, warmup=0.002, degree=0.5):
    if x < warmup:
        return x/warmup
    return (1.0 - x)**degree


SCHEDULES = {
    'warmup_cosine': warmup_cosine,
    'warmup_constant': warmup_constant,
    'warmup_linear': warmup_linear,
    'warmup_poly': warmup_poly,
}


class BertAdam(Optimizer):
    """Implements BERT version of Adam algorithm with weight decay fix.
    Params:
        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. Default: -1
        schedule: schedule to use for the warmup (see above). 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
    """
    def __init__(self, params, lr=required, warmup=-1, t_total=-1, schedule='warmup_linear',
                 b1=0.9, b2=0.999, e=1e-6, weight_decay=0.01, max_grad_norm=1.0, lars=False, weight_scaling=False):
        if lr is not required and lr < 0.0:
            raise ValueError("Invalid learning rate: {} - should be >= 0.0".format(lr))
        if schedule not in SCHEDULES:
            raise ValueError("Invalid schedule parameter: {}".format(schedule))
        if not 0.0 <= warmup < 1.0 and not warmup == -1:
            raise ValueError("Invalid warmup: {} - should be in [0.0, 1.0[ or -1".format(warmup))
        if not 0.0 <= b1 < 1.0:
            raise ValueError("Invalid b1 parameter: {} - should be in [0.0, 1.0[".format(b1))
        if not 0.0 <= b2 < 1.0:
            raise ValueError("Invalid b2 parameter: {} - should be in [0.0, 1.0[".format(b2))
        if not e >= 0.0:
            raise ValueError("Invalid epsilon value: {} - should be >= 0.0".format(e))
        defaults = dict(lr=lr, schedule=schedule, warmup=warmup, t_total=t_total,
                        b1=b1, b2=b2, e=e, weight_decay=weight_decay,
                        max_grad_norm=max_grad_norm, lars=lars, weight_scaling=weight_scaling)

        super(BertAdam, self).__init__(params, defaults)

    def get_lr(self):
        lr = []
        for group in self.param_groups:
            for p in group['params']:
                state = self.state[p]
                if len(state) == 0:
                    return [0]
                if group['t_total'] != -1:
                    schedule_fct = SCHEDULES[group['schedule']]
                    lr_scheduled = group['lr'] * schedule_fct(state['step']/group['t_total'], group['warmup'])
                else:
                    lr_scheduled = group['lr']
                lr.append(lr_scheduled)
        return lr

    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:
            if group['weight_scaling'] and 'p_norm' not in group:
                group['p_norm'] = math.sqrt(sum(torch.square(p.data).sum() for p in group['params']))
            for p in group['params']:
                if p.grad is None:
                    continue

                state = self.state[p]
                # State initialization.
                if len(state) == 0:
                    state['step'] = 0
                    if group['b1'] > 0:
                        # Exponential moving average of gradient values.
                        state['next_m'] = torch.zeros_like(p.data)
                    if group['b2'] > 0:
                        # Exponential moving average of squared gradient values.
                        state['next_v'] = torch.zeros_like(p.data)

                beta1, beta2 = group['b1'], group['b2']

                # Add grad clipping.
                if group['max_grad_norm'] > 0:
                    clip_grad_norm_(p, group['max_grad_norm'])

                # Decay the first and second moment running average coefficient.
                # In-place operations to update the averages at the same time.
                if beta1 > 0:
                    next_m = state['next_m']
                    next_m.mul_(beta1).add_(p.grad, alpha=1 - beta1)
                else:
                    next_m = p.grad
                if beta2 == -1:
                    update = next_m
                else:
                    if beta2 > 0:
                        next_v = state['next_v']
                        next_v.mul_(beta2).addcmul_(p.grad, p.grad, value=1 - beta2)
                    else:
                        next_v = torch.square(p.grad)
                    update = next_m / (next_v.sqrt() + group['e'])

                # 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.
                if group['weight_decay'] > 0.0:
                    update += group['weight_decay'] * p.data

                if group['lars']:
                    weight_norm = torch.norm(p.data).clamp(0, 10)
                    update_norm = torch.norm(update)
                    if weight_norm == 0 or update_norm == 0:
                        trust_ratio = 1
                    else:
                        trust_ratio = weight_norm / update_norm
                else:
                    trust_ratio = 1.
                state['trust_ratio'] = trust_ratio

                if group['t_total'] != -1:
                    schedule_fct = SCHEDULES[group['schedule']]
                    lr_scheduled = group['lr'] * schedule_fct(state['step']/group['t_total'], group['warmup'])
                else:
                    lr_scheduled = group['lr']

                p.data.add_(update, alpha=-lr_scheduled * trust_ratio)

                state['step'] += 1

            if group['weight_scaling']:
                p_norm = math.sqrt(sum(torch.square(p.data).sum() for p in group['params']))
                if p_norm > 0:
                    for p in group['params']:
                        p.data.mul_(group['p_norm']/p_norm)

        return loss