learning_rate.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Sat Sep  8 15:57:47 2018

@author: ai
"""

# learning rate schduler
import numpy as np
from torch.optim import Optimizer
import torch
import logging
import math
import torch.optim.lr_scheduler
from torch.optim.lr_scheduler import _LRScheduler

logger = logging.getLogger(__name__) 
class CyclicLR(object):
    """Sets the learning rate of each parameter group according to
    cyclical learning rate policy (CLR). The policy cycles the learning
    rate between two boundaries with a constant frequency, as detailed in
    the paper `Cyclical Learning Rates for Training Neural Networks`_.
    The distance between the two boundaries can be scaled on a per-iteration
    or per-cycle basis.
    Cyclical learning rate policy changes the learning rate after every batch.
    `step` should be called after a batch has been used for training.
    To resume training, save `last_batch_iteration` and use it to instantiate `CycleLR`.
    This class has three built-in policies, as put forth in the paper:
    "triangular":
        A basic triangular cycle w/ no amplitude scaling.
    "triangular2":
        A basic triangular cycle that scales initial amplitude by half each cycle.
    "exp_range":
        A cycle that scales initial amplitude by gamma**(cycle iterations) at each
        cycle iteration.
    This implementation was adapted from the github repo: `bckenstler/CLR`_
    Args:
        optimizer (Optimizer): Wrapped optimizer.
        base_lr (float or list): Initial learning rate which is the
            lower boundary in the cycle for eachparam groups.
            Default: 0.001
        max_lr (float or list): Upper boundaries in the cycle for
            each parameter group. Functionally,
            it defines the cycle amplitude (max_lr - base_lr).
            The lr at any cycle is the sum of base_lr
            and some scaling of the amplitude; therefore
            max_lr may not actually be reached depending on
            scaling function. Default: 0.006
        step_size (int): Number of training iterations per
            half cycle. Authors suggest setting step_size
            2-8 x training iterations in epoch. Default: 2000
        mode (str): One of {triangular, triangular2, exp_range}.
            Values correspond to policies detailed above.
            If scale_fn is not None, this argument is ignored.
            Default: 'triangular'
        gamma (float): Constant in 'exp_range' scaling function:
            gamma**(cycle iterations)
            Default: 1.0
        scale_fn (function): Custom scaling policy defined by a single
            argument lambda function, where
            0 <= scale_fn(x) <= 1 for all x >= 0.
            mode paramater is ignored
            Default: None
        scale_mode (str): {'cycle', 'iterations'}.
            Defines whether scale_fn is evaluated on
            cycle number or cycle iterations (training
            iterations since start of cycle).
            Default: 'cycle'
        last_batch_iteration (int): The index of the last batch. Default: -1
    Example:
        >>> optimizer = torch.optim.SGD(model.parameters(), lr=0.1, momentum=0.9)
        >>> scheduler = torch.optim.CyclicLR(optimizer)
        >>> data_loader = torch.utils.data.DataLoader(...)
        >>> for epoch in range(10):
        >>>     for batch in data_loader:
        >>>         scheduler.step()
        >>>         train_batch(...)
    .. _Cyclical Learning Rates for Training Neural Networks: https://arxiv.org/abs/1506.01186
    .. _bckenstler/CLR: https://github.com/bckenstler/CLR
    """

    def __init__(self, optimizer, base_lr=1e-3, max_lr=6e-3,
                 step_size=2000, mode='triangular', gamma=1.,
                 scale_fn=None, scale_mode='cycle', last_batch_iteration=-1):

        if not isinstance(optimizer, Optimizer):
            raise TypeError('{} is not an Optimizer'.format(
                type(optimizer).__name__))
        self.optimizer = optimizer

        if isinstance(base_lr, list) or isinstance(base_lr, tuple):
            if len(base_lr) != len(optimizer.param_groups):
                raise ValueError("expected {} base_lr, got {}".format(
                    len(optimizer.param_groups), len(base_lr)))
            self.base_lrs = list(base_lr)
        else:
            self.base_lrs = [base_lr] * len(optimizer.param_groups)

        if isinstance(max_lr, list) or isinstance(max_lr, tuple):
            if len(max_lr) != len(optimizer.param_groups):
                raise ValueError("expected {} max_lr, got {}".format(
                    len(optimizer.param_groups), len(max_lr)))
            self.max_lrs = list(max_lr)
        else:
            self.max_lrs = [max_lr] * len(optimizer.param_groups)

        self.step_size = step_size

        if mode not in ['triangular', 'triangular2', 'exp_range'] \
                and scale_fn is None:
            raise ValueError('mode is invalid and scale_fn is None')

        self.mode = mode
        self.gamma = gamma

        if scale_fn is None:
            if self.mode == 'triangular':
                self.scale_fn = self._triangular_scale_fn
                self.scale_mode = 'cycle'
            elif self.mode == 'triangular2':
                self.scale_fn = self._triangular2_scale_fn
                self.scale_mode = 'cycle'
            elif self.mode == 'exp_range':
                self.scale_fn = self._exp_range_scale_fn
                self.scale_mode = 'iterations'
        else:
            self.scale_fn = scale_fn
            self.scale_mode = scale_mode

        self.step(last_batch_iteration + 1)
        self.last_batch_iteration = last_batch_iteration

    def step(self, batch_iteration=None):
        if batch_iteration is None:
            batch_iteration = self.last_batch_iteration + 1
        self.last_batch_iteration = batch_iteration
        for param_group, lr in zip(self.optimizer.param_groups, self.get_lr()):
            param_group['lr'] = lr

    @staticmethod
    def _triangular_scale_fn(x):
        return 1.

    @staticmethod
    def _triangular2_scale_fn(x):
        return 1 / (2. ** (x - 1))

    def _exp_range_scale_fn(self, x):
        return self.gamma**(x)

    def get_lr(self):
        step_size = float(self.step_size)
        cycle = np.floor(1 + self.last_batch_iteration / (2 * step_size))
        x = np.abs(self.last_batch_iteration / step_size - 2 * cycle + 1)

        lrs = []
        param_lrs = zip(self.optimizer.param_groups, self.base_lrs, self.max_lrs)
        for param_group, base_lr, max_lr in param_lrs:
            base_height = (max_lr - base_lr) * np.maximum(0, (1 - x))
            if self.scale_mode == 'cycle':
                lr = base_lr + base_height * self.scale_fn(cycle)
            else:
                lr = base_lr + base_height * self.scale_fn(self.last_batch_iteration)
            lrs.append(lr)
        return lrs

class CosineAnnealingLR_with_Restart(_LRScheduler):
    """Set the learning rate of each parameter group using a cosine annealing
    schedule, where :math:`\eta_{max}` is set to the initial lr and
    :math:`T_{cur}` is the number of epochs since the last restart in SGDR:

    .. math::

        \eta_t = \eta_{min} + \frac{1}{2}(\eta_{max} - \eta_{min})(1 +
        \cos(\frac{T_{cur}}{T_{max}}\pi))

    When last_epoch=-1, sets initial lr as lr.

    It has been proposed in
    `SGDR: Stochastic Gradient Descent with Warm Restarts`_. The original pytorch
    implementation only implements the cosine annealing part of SGDR,
    I added my own implementation of the restarts part.
    
    Args:
        optimizer (Optimizer): Wrapped optimizer.
        T_max (int): Maximum number of iterations.
        T_mult (float): Increase T_max by a factor of T_mult
        eta_min (float): Minimum learning rate. Default: 0.
        last_epoch (int): The index of last epoch. Default: -1.
        model (pytorch model): The model to save.
        out_dir (str): Directory to save snapshots
        take_snapshot (bool): Whether to save snapshots at every restart

    .. _SGDR\: Stochastic Gradient Descent with Warm Restarts:
        https://arxiv.org/abs/1608.03983
    """

    def __init__(self, optimizer, T_max, T_mult, model, out_dir, take_snapshot, eta_min=0, last_epoch=-1):
        self.T_max = T_max
        self.T_mult = T_mult
        self.Te = self.T_max
        self.eta_min = eta_min
        self.current_epoch = last_epoch
        
        self.model = model
        self.out_dir = out_dir
        self.take_snapshot = take_snapshot
        
        self.lr_history = []
        
        super(CosineAnnealingLR_with_Restart, self).__init__(optimizer, last_epoch)

    def get_lr(self):
        new_lrs = [self.eta_min + (base_lr - self.eta_min) *
                (1 + math.cos(math.pi * self.current_epoch / self.Te)) / 2

                for base_lr in self.base_lrs]
        
        self.lr_history.append(new_lrs)
        return new_lrs
    
    def step(self, epoch=None):
        if epoch is None:
        
            epoch = self.last_epoch + 1
        self.last_epoch = epoch
        self.current_epoch += 1
        
        for param_group, lr in zip(self.optimizer.param_groups, self.get_lr()):
            param_group['lr'] = lr
        
        ## restart
        if self.current_epoch == self.Te:
            print("restart at epoch {:03d}".format(self.last_epoch + 1))
            
            if self.take_snapshot:
                torch.save({
                    'epoch': self.T_max,
                    'state_dict': self.model.state_dict()
                }, self.out_dir + "/" + 'snapshot_e_{:03d}.pth.tar'.format(self.T_max))
            
            ## reset epochs since the last reset
            self.current_epoch = 0
            
            ## reset the next goal
            self.Te = int(self.Te * self.T_mult)
            self.T_max = self.T_max + self.Te