Source code for deeppavlov.core.models.lr_scheduled_model

# Copyright 2017 Neural Networks and Deep Learning lab, MIPT
#
# 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.

from typing import Any, Union, Tuple, List, Optional
from logging import getLogger
from abc import abstractmethod
import math
from enum import IntEnum

import numpy as np

from deeppavlov.core.common.errors import ConfigError


log = getLogger(__name__)


class DecayType(IntEnum):
    """ Data class, each decay type is assigned a number. """
    NO = 1
    LINEAR = 2
    COSINE = 3
    EXPONENTIAL = 4
    POLYNOMIAL = 5
    ONECYCLE = 6
    TRAPEZOID = 7

    @classmethod
    def from_str(cls, label: str) -> int:
        """
        Convert given string label of decay type to special index

        Args:
            label: name of decay type.
                Set of values: `"linear"`, `"cosine"`, `"exponential"`,
                `"onecycle"`, `"trapezoid"`, `["polynomial", K]`, where K is a polynomial power

        Returns:
            index of decay type
        """
        label_norm = label.replace('1', 'one').upper()
        if label_norm in cls.__members__:
            return DecayType[label_norm]
        else:
            raise NotImplementedError


class DecayScheduler:
    """
    Given initial and endvalue, this class generates the next value
    depending on decay type and number of iterations. (by calling next_val().)
    """

    def __init__(self, dec_type: Union[str, DecayType], start_val: float,
                 num_it: int = 0, end_val: float = None, extra: float = None) -> None:
        if isinstance(dec_type, DecayType):
            self.dec_type = dec_type
        else:
            self.dec_type = DecayType.from_str(dec_type)
        self.nb, self.extra = num_it, extra
        self.start_val, self.end_val = start_val, end_val
        self.iters = 0
        if self.end_val is None and not (self.dec_type in [1, 4]):
            self.end_val = 0
        if self.dec_type == DecayType.ONECYCLE:
            self.cycle_nb = math.ceil(self.nb / 2)
            self.div = 1.0 if not self.start_val else self.end_val / self.start_val
        if self.dec_type == DecayType.TRAPEZOID:
            self.div = 1.0 if not self.start_val else self.end_val / self.start_val

    def __str__(self):
        return f"DecayScheduler(start_val={self.start_val}, end_val={self.end_val}"\
            f", dec_type={self.dec_type.name}, num_it={self.nb}, extra={self.extra})"

    def next_val(self) -> float:
        self.iters = min(self.iters + 1, self.nb)
        # print(f"iters = {self.iters}/{self.nb}")
        if self.dec_type == DecayType.NO:
            return self.start_val
        elif self.dec_type == DecayType.LINEAR:
            pct = self.iters / self.nb
            return self.start_val + pct * (self.end_val - self.start_val)
        elif self.dec_type == DecayType.COSINE:
            cos_out = math.cos(math.pi * self.iters / self.nb) + 1
            return self.end_val + (self.start_val - self.end_val) / 2 * cos_out
        elif self.dec_type == DecayType.EXPONENTIAL:
            ratio = self.end_val / self.start_val
            return self.start_val * (ratio ** (self.iters / self.nb))
        elif self.dec_type == DecayType.POLYNOMIAL:
            delta_val = self.start_val - self.end_val
            return self.end_val + delta_val * (1 - self.iters / self.nb) ** self.extra
        elif self.dec_type == DecayType.ONECYCLE:
            if self.iters > self.cycle_nb:
                # decaying from end_val to start_val for cycle_nb steps
                pct = 1 - (self.iters - self.cycle_nb) / self.cycle_nb
                return self.start_val * (1 + pct * (self.div - 1))
            else:
                # raising from start_val to end_val for cycle_nb steps
                pct = self.iters / self.cycle_nb
                return self.start_val * (1 + pct * (self.div - 1))
        elif self.dec_type == DecayType.TRAPEZOID:
            if self.iters > 0.6 * self.nb:
                # decaying from end_val to start_val for 4/10 * nb steps
                pct = 2.5 * (self.nb - self.iters) / self.nb
                return self.start_val * (1 + pct * (self.div - 1))
            elif self.iters > 0.1 * self.nb:
                # constant end_val
                return self.end_val
            else:
                # raising from start_val to end_val for 1/10 * nb steps
                pct = 10.0 * self.iters / self.nb
                return self.start_val * (1 + pct * (self.div - 1))


DType = Union[str, DecayType]


[docs]class LRScheduledModel: """ Abstract model enhanced with optimizer, learning rate and momentum management and search. Args: learning_rate: learning rate value or ranges learning_rate_decay: learning rate decay type. Set of values: `"linear"`, `"onecycle"`, `"trapezoid"`, `"exponential"`, `"cosine"`, `["polynomial", K]`, where K is a polynomial power learning_rate_decay_epochs: number of epochs for learning rate decay process learning_rate_decay_batches: number of batches for learning rate decay process learning_rate_drop_div: division coefficient for learning rate in case of exceeding patience `learning_rate_drop_patience` learning_rate_drop_patience: patience limit of loss increase momentum: range of momentum values momentum_decay: momentum decay type. Set of values: `"linear"`, `"onecycle"`, `"trapezoid"`, `"exponential"`, `"cosine"`, `["polynomial", K]`, where K is a polynomial power momentum_decay_epochs: number of epochs for momentum decay process momentum_decay_batches: number of batches for momentum decay process fit_batch_size: batch size when fitting learning rate fit_learning_rate: range of learning rate values to explore fit_learning_rate_div: division coefficient for best learning rate obtained from fitting, divided learning rate value will be used when training model fit_beta: smoothing coefficient for loss calculation when fitting learning rate fit_min_batches: number of batches to train model on before fitting learning rate fit_max_batches: number of batches to train model on when fitting learning rate load_before_drop: set True to load saved model from disk when learning rate is dropped, set False to continue training current model *args: other parameters **kwargs: other parameters """ @abstractmethod def _init_learning_rate_variable(self): pass @abstractmethod def _init_momentum_variable(self): pass @abstractmethod def _update_graph_variables(self, learning_rate=None, momentum=None): """ Update learning rate in graph if `learning_rate` is not None, Update momentum in graph if `momentum` is not None """ if learning_rate is not None: # update learning rate pass if momentum is not None: # update momentum pass def __init__(self, learning_rate: Union[None, float, Tuple[float, float]] = None, learning_rate_decay: Union[DType, Tuple[DType, float]] = DecayType.NO, learning_rate_decay_epochs: int = 0, learning_rate_decay_batches: int = 0, learning_rate_drop_div: float = 2.0, learning_rate_drop_patience: Optional[int] = None, momentum: Union[None, float, Tuple[float, float]] = None, momentum_decay: Union[DType, Tuple[DType, float]] = DecayType.NO, momentum_decay_epochs: int = 0, momentum_decay_batches: int = 0, fit_batch_size: Union[None, int, str] = None, fit_learning_rate: Tuple[float, float] = (1e-7, 100), fit_learning_rate_div: float = 10., fit_beta: float = 0.98, fit_min_batches: int = 10, fit_max_batches: Optional[int] = None, load_before_drop: bool = False, *args, **kwargs) -> None: """ Initialize learning rate scheduler """ if learning_rate_decay_epochs and learning_rate_decay_batches: raise ConfigError("isn't able to update learning rate every batch" " and every epoch simultaneously") if momentum_decay_epochs and momentum_decay_batches: raise ConfigError("isn't able to update momentum every batch" " and every epoch simultaneously") start_val, end_val = learning_rate, None if isinstance(learning_rate, (tuple, list)): start_val, end_val = learning_rate dec_type, extra = learning_rate_decay, None if isinstance(learning_rate_decay, (tuple, list)): dec_type, extra = learning_rate_decay self._lr = start_val num_it, self._lr_update_on_batch = learning_rate_decay_epochs, False if learning_rate_decay_batches > 0: num_it, self._lr_update_on_batch = learning_rate_decay_batches, True self._lr_schedule = DecayScheduler(start_val=start_val, end_val=end_val, num_it=num_it, dec_type=dec_type, extra=extra) self._lr_var = self._init_learning_rate_variable() start_val, end_val = momentum, None if isinstance(momentum, (tuple, list)): start_val, end_val = momentum dec_type, extra = momentum_decay, None if isinstance(momentum_decay, (tuple, list)): dec_type, extra = momentum_decay self._mom = start_val num_it, self._mom_update_on_batch = momentum_decay_epochs, False if momentum_decay_batches > 0: num_it, self._mom_update_on_batch = momentum_decay_batches, False self._mom_schedule = DecayScheduler(start_val=start_val, end_val=end_val, num_it=num_it, dec_type=dec_type, extra=extra) self._mom_var = self._init_momentum_variable() self._learning_rate_drop_patience = learning_rate_drop_patience self._learning_rate_drop_div = learning_rate_drop_div self._learning_rate_cur_impatience = 0. self._learning_rate_last_impatience = 0. self._learning_rate_cur_div = 1. self._load_before_drop = load_before_drop self._fit_batch_size = fit_batch_size self._fit_learning_rate = fit_learning_rate self._fit_learning_rate_div = fit_learning_rate_div self._fit_beta = fit_beta self._fit_min_batches = fit_min_batches self._fit_max_batches = fit_max_batches def get_learning_rate(self): """ Return current learning rate value Returns: learning rate """ if self._lr is None: raise ConfigError("Please specify `learning_rate` parameter" " before training") return self._lr def get_learning_rate_variable(self): """ Return current learning rate variable Returns: learning rate variable """ return self._lr_var def get_momentum(self): """ Return current momentum value Returns: momentum """ return self._mom def get_momentum_variable(self): """ Return current momentum variable Returns: momentum variable """ return self._mom_var def fit(self, *args): """ Find the best learning rate schedule, and set obtained values of learning rate and momentum for further model training. Best learning rate will be divided by `fit_learning_rate_div` for further training model. Args: *args: arguments Returns: """ data = list(zip(*args)) self.save() if self._fit_batch_size is None: raise ConfigError("in order to use fit() method" " set `fit_batch_size` parameter") bs = int(self._fit_batch_size) data_len = len(data) num_batches = self._fit_max_batches or ((data_len - 1) // bs + 1) avg_loss = 0. best_loss = float('inf') lrs, losses = [], [] _lr_find_schedule = DecayScheduler(start_val=self._fit_learning_rate[0], end_val=self._fit_learning_rate[1], dec_type="exponential", num_it=num_batches) self._lr = _lr_find_schedule.start_val self._mom = 0. self._update_graph_variables(learning_rate=self._lr, momentum=self._mom) best_lr = _lr_find_schedule.start_val for i in range(num_batches): batch_start = (i * bs) % data_len batch_end = batch_start + bs report = self.train_on_batch(*zip(*data[batch_start:batch_end])) if not isinstance(report, dict): report = {'loss': report} # Calculating smoothed loss avg_loss = self._fit_beta*avg_loss + (1 - self._fit_beta)*report['loss'] smoothed_loss = avg_loss / (1 - self._fit_beta**(i + 1)) lrs.append(self._lr) losses.append(smoothed_loss) log.info(f"Batch {i}/{num_batches}: smooth_loss = {smoothed_loss}" f", lr = {self._lr}, best_lr = {best_lr}") if math.isnan(smoothed_loss) or (smoothed_loss > 4 * best_loss): break if (smoothed_loss < best_loss) and (i >= self._fit_min_batches): best_loss = smoothed_loss best_lr = self._lr self._lr = _lr_find_schedule.next_val() self._update_graph_variables(learning_rate=self._lr) if i >= num_batches: break # best_lr /= 10 end_val = self._get_best(lrs, losses) start_val = end_val if self._lr_schedule.dec_type in (DecayType.ONECYCLE, DecayType.TRAPEZOID): start_val = end_val / self._fit_learning_rate_div elif self._lr_schedule.dec_type in (DecayType.POLYNOMIAL, DecayType.EXPONENTIAL, DecayType.LINEAR, DecayType.COSINE): start_val = end_val end_val = end_val / self._fit_learning_rate_div self._lr_schedule = DecayScheduler(start_val=start_val, end_val=end_val, num_it=self._lr_schedule.nb, dec_type=self._lr_schedule.dec_type, extra=self._lr_schedule.extra) log.info(f"Found best learning rate value = {best_lr}" f", setting new learning rate schedule with {self._lr_schedule}.") self.load() self._lr = self._lr_schedule.start_val self._mom = self._mom_schedule.start_val self._update_graph_variables(learning_rate=self._lr, momentum=self._mom) return {'smoothed_loss': losses, 'learning_rate': lrs} @staticmethod def _get_best(values: List[float], losses: List[float], max_loss_div: float = 0.9, min_val_div: float = 10.0) -> float: """ Find the best value according to given losses Args: values: list of considered values losses: list of obtained loss values corresponding to `values` max_loss_div: maximal divergence of loss to be considered significant min_val_div: minimum divergence of loss to be considered significant Returns: best value divided by `min_val_div` """ assert len(values) == len(losses), "lengths of values and losses should be equal" min_ind = np.argmin(losses) for i in range(min_ind - 1, 0, -1): if (losses[i] * max_loss_div > losses[min_ind]) or\ (values[i] * min_val_div < values[min_ind]): return values[i + 1] return values[min_ind] / min_val_div def process_event(self, event_name: str, data: dict) -> None: """ Update learning rate and momentum variables after event (given by `event_name`) Args: event_name: name of event after which the method was called. Set of values: `"after_validation"`, `"after_batch"`, `"after_epoch"`, `"after_train_log"` data: dictionary with parameters values Returns: None """ if event_name == "after_validation": if data['impatience'] > self._learning_rate_last_impatience: self._learning_rate_cur_impatience += 1 else: self._learning_rate_cur_impatience = 0 self._learning_rate_last_impatience = data['impatience'] if (self._learning_rate_drop_patience is not None) and\ (self._learning_rate_cur_impatience >= self._learning_rate_drop_patience): self._learning_rate_cur_impatience = 0 self._learning_rate_cur_div *= self._learning_rate_drop_div self._lr /= self._learning_rate_drop_div if self._load_before_drop: self.load(path=self.save_path) self._update_graph_variables(momentum=self._mom) self._update_graph_variables(learning_rate=self._lr) log.info(f"New learning rate dividor = {self._learning_rate_cur_div}") if event_name == 'after_batch': if (self._lr is not None) and self._lr_update_on_batch: self._lr = self._lr_schedule.next_val() / self._learning_rate_cur_div self._update_graph_variables(learning_rate=self._lr) if (self._mom is not None) and self._mom_update_on_batch: self._mom = min(1., max(0., self._mom_schedule.next_val())) self._update_graph_variables(momentum=self._mom) if event_name == 'after_epoch': if (self._lr is not None) and not self._lr_update_on_batch: self._lr = self._lr_schedule.next_val() / self._learning_rate_cur_div self._update_graph_variables(learning_rate=self._lr) if (self._mom is not None) and not self._mom_update_on_batch: self._mom = min(1., max(0., self._mom_schedule.next_val())) self._update_graph_variables(momentum=self._mom) if event_name == 'after_train_log': if (self._lr is not None) and ('learning_rate' not in data): data['learning_rate'] = self._lr if (self._mom is not None) and ('momentum' not in data): data['momentum'] = self._mom