mindspore/example/resnet50_imagenet2012_THOR/lr_generator.py

# Copyright 2020 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""learning rate generator"""
import math

import numpy as np


def linear_warmup_lr(current_step, warmup_steps, base_lr, init_lr):
    """linear_warmup_lr"""
    lr_inc = (float(base_lr) - float(init_lr)) / float(warmup_steps)
    lr = float(init_lr) + lr_inc * current_step
    return lr


def cosine_annealing_lr(lr, steps_per_epoch, warmup_epochs, max_epoch, T_max, eta_min=0, num_periods=0.5):
    """linear_warmup_lr"""
    base_lr = lr
    warmup_init_lr = 0
    total_steps = int(max_epoch * steps_per_epoch)
    warmup_steps = int(warmup_epochs * steps_per_epoch)
    decay_steps = total_steps - warmup_steps
    lr_each_step = []
    for i in range(total_steps):
        if i < warmup_steps:
            lr = linear_warmup_lr(i + 1, warmup_steps, base_lr, warmup_init_lr)
        else:
            # linear_decay = (total_steps - i) / decay_steps
            cosine_decay = 0.5 * (1 + math.cos(math.pi * i / decay_steps))
            decayed = cosine_decay
            lr = base_lr * decayed
        lr_each_step.append(lr)
    return np.array(lr_each_step).astype(np.float32)


def warmup_cosine_annealing_lr(lr, steps_per_epoch, warmup_epochs, max_epoch, T_max, eta_min=0, num_periods=0.5):
    """warmup_cosine_annealing_lr"""
    base_lr = lr
    warmup_init_lr = 0
    total_steps = int(max_epoch * steps_per_epoch * 0.99)
    warmup_steps = int(warmup_epochs * steps_per_epoch)
    decay_steps = total_steps - warmup_steps
    lr_each_step = []
    for i in range(total_steps):
        if i < warmup_steps:
            lr = linear_warmup_lr(i + 1, warmup_steps, base_lr, warmup_init_lr)
        else:
            linear_decay = (total_steps - i) / decay_steps
            cosine_decay = 0.5 * (1 + math.cos(math.pi * 2 * num_periods * i / decay_steps))
            decayed = linear_decay * cosine_decay
            lr = base_lr * decayed + 0.000005
        lr_each_step.append(lr)
    return np.array(lr_each_step).astype(np.float32)


def get_lr(global_step, lr_init, lr_end, lr_max, warmup_epochs, total_epochs, steps_per_epoch, lr_decay_mode):
    """
    generate learning rate array

    Args:
       global_step(int): total steps of the training
       lr_init(float): init learning rate
       lr_end(float): end learning rate
       lr_max(float): max learning rate
       warmup_epochs(int): number of warmup epochs
       total_epochs(int): total epoch of training
       steps_per_epoch(int): steps of one epoch
       lr_decay_mode(string): learning rate decay mode, including steps, poly or default

    Returns:
       np.array, learning rate array
    """
    lr_each_step = []
    total_steps = steps_per_epoch * total_epochs
    warmup_steps = steps_per_epoch * warmup_epochs
    if lr_decay_mode == 'steps':
        decay_epoch_index = [0.3 * total_steps, 0.6 * total_steps, 0.8 * total_steps]
        for i in range(total_steps):
            if i < decay_epoch_index[0]:
                lr = lr_max
            elif i < decay_epoch_index[1]:
                lr = lr_max * 0.1
            elif i < decay_epoch_index[2]:
                lr = lr_max * 0.01
            else:
                lr = lr_max * 0.001
            lr_each_step.append(lr)
    elif lr_decay_mode == 'poly':
        if warmup_steps != 0:
            inc_each_step = (float(lr_max) - float(lr_init)) / float(warmup_steps)
        else:
            inc_each_step = 0
        for i in range(total_steps):
            if i < warmup_steps:
                lr = float(lr_init) + inc_each_step * float(i)
            else:
                base = (1.0 - (float(i) - float(warmup_steps)) / (float(total_steps) - float(warmup_steps)))
                lr = float(lr_max) * base * base
                if lr < 0.0:
                    lr = 0.0
            lr_each_step.append(lr)
    else:
        for i in range(total_steps):
            if i < warmup_steps:
                lr = lr_init + (lr_max - lr_init) * i / warmup_steps
            else:
                lr = lr_max - (lr_max - lr_end) * (i - warmup_steps) / (total_steps - warmup_steps)
            lr_each_step.append(lr)

    current_step = global_step
    lr_each_step = np.array(lr_each_step).astype(np.float32)
    learning_rate = lr_each_step[current_step:]

    return learning_rate