optimization_openai.py

# coding=utf-8
# Copyright 2018 The Open AI 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 OpenAI GPT model."""

import math
import torch
from torch.optim import Optimizer
from torch.optim.optimizer import required
from torch.nn.utils import clip_grad_norm_
import logging
from .optimization import *

logger = logging.getLogger(__name__)


class OpenAIAdam(Optimizer):
    """Implements Open AI version of Adam algorithm with weight decay fix.
    """
    def __init__(self, params, lr=required, schedule='warmup_linear', warmup=-1, t_total=-1,
                 b1=0.9, b2=0.999, e=1e-8, weight_decay=0,
                 vector_l2=False, max_grad_norm=-1, **kwargs):
        if lr is not required and lr < 0.0:
            raise ValueError("Invalid learning rate: {} - should be >= 0.0".format(lr))
        if not isinstance(schedule, LRSchedule) and schedule not in SCHEDULES:
            raise ValueError("Invalid schedule parameter: {}".format(schedule))
        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))
        # initialize schedule object
        if not isinstance(schedule, LRSchedule):
            schedule_type = SCHEDULES[schedule]
            schedule = schedule_type(warmup=warmup, t_total=t_total)
        else:
            if warmup != -1 or t_total != -1:
                logger.warning("Non-default warmup and t_total are ineffective when LRSchedule object is provided. "
                               "Please specify custom warmup and t_total in LRSchedule object.")
        defaults = dict(lr=lr, schedule=schedule,
                        b1=b1, b2=b2, e=e, weight_decay=weight_decay, vector_l2=vector_l2,
                        max_grad_norm=max_grad_norm)
        super(OpenAIAdam, 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]
                lr_scheduled = group['lr']
                lr_scheduled *= group['schedule'].get_lr(state['step'])
                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:
            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['b1'], group['b2']

                state['step'] += 1

                # 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
                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['e'])

                bias_correction1 = 1 - beta1 ** state['step']
                bias_correction2 = 1 - beta2 ** state['step']

                lr_scheduled = group['lr']
                lr_scheduled *= group['schedule'].get_lr(state['step'])

                step_size = lr_scheduled * math.sqrt(bias_correction2) / bias_correction1

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

                # Add weight decay at the end (fixed version)
                if (len(p.size()) > 1 or group['vector_l2']) and group['weight_decay'] > 0:
                    p.data.add_(-lr_scheduled * group['weight_decay'], p.data)

        return loss