learning_rate_decay.py 8.3 KB
Newer Older
Q
Qiao Longfei 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
# Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved
#
# 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.

import layers
from framework import Variable
17
from initializer import init_on_cpu
Q
Qiao Longfei 已提交
18

19 20 21 22
__all__ = [
    'exponential_decay', 'natural_exp_decay', 'inverse_time_decay',
    'polynomial_decay', 'piecewise_decay'
]
Q
Qiao Longfei 已提交
23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57
"""
When training a model, it's often useful to decay the
learning rate during training process, this is called
learning_rate_decay. There are many strategies to do
this, this module will provide some classical method.
User can also implement their own learning_rate_decay
strategy according to this module.
"""


def exponential_decay(learning_rate,
                      global_step,
                      decay_steps,
                      decay_rate,
                      staircase=False):
    """Applies exponential decay to the learning rate.

    ```python
    decayed_learning_rate = learning_rate *
            decay_rate ^ (global_step / decay_steps)
    ```
    Args:
        learning_rate: A scalar float32 value or a Variable. This
          will be the initial learning rate during training
        global_step: A Variable that record the training step.
        decay_steps: A Python `int32` number.
        decay_rate: A Python `float` number.
        staircase: Boolean. If set true, decay the learning rate every decay_steps.

    Returns:
        The decayed learning rate
    """
    if not isinstance(global_step, Variable):
        raise ValueError("global_step is required for exponential_decay.")

58 59 60 61 62 63 64 65
    with init_on_cpu():
        # update learning_rate
        div_res = global_step / decay_steps
        if staircase:
            div_res = layers.floor(x=div_res)
        decayed_lr = learning_rate * (decay_rate**div_res)

    return decayed_lr
Q
Qiao Longfei 已提交
66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94


def natural_exp_decay(learning_rate,
                      global_step,
                      decay_steps,
                      decay_rate,
                      staircase=False):
    """Applies natural exponential decay to the initial learning rate.

    ```python
    if not staircase:
        decayed_learning_rate = learning_rate * exp(- decay_rate * (global_step / decay_steps))
    else:
        decayed_learning_rate = learning_rate * exp(- decay_rate * (global_step / decay_steps))
    ```
    Args:
        learning_rate: A scalar float32 value or a Variable. This
          will be the initial learning rate during training
        global_step: A Variable that record the training step.
        decay_steps: A Python `int32` number.
        decay_rate: A Python `float` number.
        staircase: Boolean. If set true, decay the learning rate every decay_steps.

    Returns:
        The decayed learning rate
    """
    if not isinstance(global_step, Variable):
        raise ValueError("global_step is required for natural_exp_decay.")

95 96 97 98 99 100 101
    with init_on_cpu():
        div_res = global_step / decay_steps
        if staircase:
            div_res = layers.floor(x=div_res)
        decayed_lr = learning_rate * layers.exp(x=(-1 * decay_rate * div_res))

    return decayed_lr
Q
Qiao Longfei 已提交
102 103 104 105 106 107 108 109 110 111 112 113


def inverse_time_decay(learning_rate,
                       global_step,
                       decay_steps,
                       decay_rate,
                       staircase=False):
    """Applies inverse time decay to the initial learning rate.

    ```python
    if staircase:
      decayed_learning_rate = learning_rate / (1 + decay_rate * floor(global_step / decay_step))
114
    else:
Q
Qiao Longfei 已提交
115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130
      decayed_learning_rate = learning_rate / (1 + decay_rate * global_step / decay_step)
    ```
    Args:
        learning_rate: A scalar float32 value or a Variable. This
          will be the initial learning rate during training
        global_step: A Variable that record the training step.
        decay_steps: A Python `int32` number.
        decay_rate: A Python `float` number.
        staircase: Boolean. If set true, decay the learning rate every decay_steps.

    Returns:
        The decayed learning rate
    """
    if not isinstance(global_step, Variable):
        raise ValueError("global_step is required for inverse_time_decay.")

131 132 133 134 135 136
    with init_on_cpu():
        div_res = global_step / decay_steps
        if staircase:
            div_res = layers.floor(x=div_res)

        decayed_lr = learning_rate / (1 + decay_rate * div_res)
Q
Qiao Longfei 已提交
137

138
    return decayed_lr
139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172


def polynomial_decay(learning_rate,
                     global_step,
                     decay_steps,
                     end_learning_rate=0.0001,
                     power=1.0,
                     cycle=False):
    """Applies polynomial decay to the initial learning rate.

    ```python
    if cycle:
        decay_steps = decay_steps * ceil(global_step / decay_steps)
    else:
        global_step = min(global_step, decay_steps)
    decayed_learning_rate = (learning_rate - end_learning_rate) *
                      (1 - global_step / decay_steps) ^ power +
                      end_learning_rate
    ```
    Args:
        learning_rate: A scalar float32 value or a Variable. This
          will be the initial learning rate during training
        global_step: A Variable that record the training step.
        decay_steps: A Python `int32` number.
        end_learning_rate: A Python `float` number.
        power: A Python `float` number
        cycle: Boolean. If set true, decay the learning rate every decay_steps.

    Returns:
        The decayed learning rate
    """
    if not isinstance(global_step, Variable):
        raise ValueError("global_step is required for inverse_time_decay.")

173 174 175 176 177 178 179 180 181
    with init_on_cpu():
        if cycle:
            div_res = layers.ceil(x=(global_step / decay_steps))
            zero_var = layers.fill_constant(
                shape=[1], dtype='float32', value=0.0)
            one_var = layers.fill_constant(
                shape=[1], dtype='float32', value=1.0)

            with layers.Switch() as switch:
182
                with switch.case(global_step == zero_var):
183 184 185 186 187 188 189 190 191 192 193
                    layers.assign(input=one_var, output=div_res)
            decay_steps = decay_steps * div_res
        else:
            decay_steps_var = layers.fill_constant(
                shape=[1], dtype='float32', value=float(decay_steps))
            global_step = layers.elementwise_min(
                x=global_step, y=decay_steps_var)

        decayed_lr = (learning_rate - end_learning_rate) * \
                     ((1 - global_step / decay_steps) ** power) + end_learning_rate
    return decayed_lr
194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217


def piecewise_decay(global_step, boundaries, values):
    """Applies piecewise decay to the initial learning rate.

    ```python
    boundaries = [10000, 20000]
    values = [1.0, 0.5, 0.1]

    if step < 10000:
        learning_rate = 1.0
    elif step >= 10000 and step < 20000:
        learning_rate = 0.5
    else:
        learning_rate = 0.1
    ```
    """

    if len(values) - len(boundaries) != 1:
        raise ValueError("len(values) - len(boundaries) should be 1")

    if not isinstance(global_step, Variable):
        raise ValueError("global_step is required for piecewise_decay.")

218 219 220 221 222 223 224 225 226 227 228 229 230 231
    with init_on_cpu():
        lr = layers.create_global_var(
            shape=[1],
            value=0.0,
            dtype='float32',
            persistable=True,
            name="learning_rate")

        with layers.Switch() as switch:
            for i in range(len(boundaries)):
                boundary_val = layers.fill_constant(
                    shape=[1], dtype='float32', value=float(boundaries[i]))
                value_var = layers.fill_constant(
                    shape=[1], dtype='float32', value=float(values[i]))
232
                with switch.case(global_step < boundary_val):
233 234 235 236 237 238 239
                    layers.assign(value_var, lr)
            last_value_var = layers.fill_constant(
                shape=[1],
                dtype='float32',
                value=float(values[len(values) - 1]))
            with switch.default():
                layers.assign(last_value_var, lr)
240 241

    return lr