learning_rate_scheduler.py 9.8 KB
Newer Older
Q
Qiao Longfei 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13
# 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.
Y
yuyang18 已提交
14 15 16 17 18 19 20 21
"""
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.
"""
Q
Qiao Longfei 已提交
22

23 24 25 26 27
import control_flow
import nn
import ops
import tensor
from ..initializer import init_on_cpu
Q
Qiao Longfei 已提交
28

29 30
__all__ = [
    'exponential_decay', 'natural_exp_decay', 'inverse_time_decay',
31
    'polynomial_decay', 'piecewise_decay', 'noam_decay'
32
]
Q
Qiao Longfei 已提交
33 34


35
def _decay_step_counter(begin=0):
Y
Yu Yang 已提交
36
    # the first global step is zero in learning rate decay
37
    global_step = nn.autoincreased_step_counter(
38
        counter_name='@LR_DECAY_COUNTER@', begin=begin, step=1)
39
    global_step = tensor.cast(global_step, 'float32')
Y
Yu Yang 已提交
40 41 42
    return global_step


43
def noam_decay(d_model, warmup_steps):
Y
yuyang18 已提交
44 45 46 47 48 49 50 51 52 53
    """
    Noam decay method. The numpy implementation of noam decay as follows.

    >>> import numpy as np
    >>> lr_value = np.power(d_model, -0.5) * np.min([
    >>>                         np.power(current_steps, -0.5),
    >>>                         np.power(warmup_steps, -1.5) * current_steps])

    Please reference `attention is all you need
    <https://arxiv.org/pdf/1706.03762.pdf>`_.
54 55 56

    Args:
        d_model(Variable): The dimensionality of input and output of model.
Y
yuyang18 已提交
57

58 59 60 61 62 63 64 65 66 67 68 69 70 71
        warmup_steps(Variable): A super parameter.

    Returns:
        The decayed learning rate.
    """
    global_step = _decay_step_counter(1)
    with init_on_cpu():
        a = global_step**-0.5
        b = (warmup_steps**-1.5) * global_step
        lr_value = (d_model**-0.5) * ops.elementwise_min(a, b)

    return lr_value


Y
Yu Yang 已提交
72
def exponential_decay(learning_rate, decay_steps, decay_rate, staircase=False):
F
fengjiayi 已提交
73 74 75 76 77 78 79 80 81 82 83
    """
    Applies exponential decay to the learning rate. 

    When training a model, it is often recommended to lower the learning rate as the 
    training progresses. By using this function, the learning rate will be decayed by 
    'decay_rate' every 'decay_steps' steps.

    >>> if staircase == True:
    >>>     decayed_learning_rate = learning_rate * decay_rate ^ floor(global_step / decay_steps)
    >>> else:
    >>>     decayed_learning_rate = learning_rate * decay_rate ^ (global_step / decay_steps)
Q
Qiao Longfei 已提交
84 85

    Args:
F
fengjiayi 已提交
86 87 88 89 90
        learning_rate(Variable|float): The initial learning rate.
        decay_steps(int): See the decay computation above.
        decay_rate(float): The decay rate. See the decay computation above.
        staircase(Boolean): If True, decay the learning rate at discrete intervals.
                            Default: False
Q
Qiao Longfei 已提交
91 92 93

    Returns:
        The decayed learning rate
F
fengjiayi 已提交
94 95 96 97 98 99 100 101 102 103 104 105 106

    Examples:
        .. code-block:: python

          base_lr = 0.1
          sgd_optimizer = fluid.optimizer.SGD(
                learning_rate=fluid.layers.exponential_decay(
                    learning_rate=base_lr,
                    decay_steps=10000,
                    decay_rate=0.5,
                    staircase=True))
          sgd_optimizer.minimize(avg_cost)

Q
Qiao Longfei 已提交
107
    """
Y
Yu Yang 已提交
108
    global_step = _decay_step_counter()
Q
Qiao Longfei 已提交
109

110 111 112 113
    with init_on_cpu():
        # update learning_rate
        div_res = global_step / decay_steps
        if staircase:
114
            div_res = ops.floor(div_res)
115 116 117
        decayed_lr = learning_rate * (decay_rate**div_res)

    return decayed_lr
Q
Qiao Longfei 已提交
118 119


Y
Yu Yang 已提交
120
def natural_exp_decay(learning_rate, decay_steps, decay_rate, staircase=False):
Q
Qiao Longfei 已提交
121 122
    """Applies natural exponential decay to the initial learning rate.

Y
Yu Yang 已提交
123 124 125 126 127
    >>> 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))

Q
Qiao Longfei 已提交
128 129 130 131 132 133 134 135 136 137
    Args:
        learning_rate: A scalar float32 value or a Variable. This
          will be the initial learning rate during training
        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
    """
Y
Yu Yang 已提交
138
    global_step = _decay_step_counter()
Q
Qiao Longfei 已提交
139

140 141 142
    with init_on_cpu():
        div_res = global_step / decay_steps
        if staircase:
143 144
            div_res = ops.floor(div_res)
        decayed_lr = learning_rate * ops.exp(-1 * decay_rate * div_res)
145 146

    return decayed_lr
Q
Qiao Longfei 已提交
147 148


Y
Yu Yang 已提交
149
def inverse_time_decay(learning_rate, decay_steps, decay_rate, staircase=False):
F
fengjiayi 已提交
150 151
    """
    Applies inverse time decay to the initial learning rate.
Q
Qiao Longfei 已提交
152

F
fengjiayi 已提交
153 154 155 156 157
    When training a model, it is often recommended to lower the learning rate as the 
    training progresses. By using this function, an inverse decay function will be 
    applied to the initial learning rate.

    >>> if staircase == True:
Y
Yu Yang 已提交
158 159 160 161
    >>>     decayed_learning_rate = learning_rate / (1 + decay_rate * floor(global_step / decay_step))
    >>> else:
    >>>     decayed_learning_rate = learning_rate / (1 + decay_rate * global_step / decay_step)

Q
Qiao Longfei 已提交
162
    Args:
F
fengjiayi 已提交
163 164 165 166 167
        learning_rate(Variable|float): The initial learning rate.
        decay_steps(int): See the decay computation above.
        decay_rate(float): The decay rate. See the decay computation above.
        staircase(Boolean): If True, decay the learning rate at discrete intervals.
                            Default: False
Q
Qiao Longfei 已提交
168 169 170

    Returns:
        The decayed learning rate
F
fengjiayi 已提交
171 172 173 174 175 176 177 178 179 180 181 182

    Examples:
        .. code-block:: python

          base_lr = 0.1
          sgd_optimizer = fluid.optimizer.SGD(
                learning_rate=fluid.layers.inverse_time_decay(
                    learning_rate=base_lr,
                    decay_steps=10000,
                    decay_rate=0.5,
                    staircase=True))
          sgd_optimizer.minimize(avg_cost)
Q
Qiao Longfei 已提交
183
    """
Y
Yu Yang 已提交
184
    global_step = _decay_step_counter()
Q
Qiao Longfei 已提交
185

186 187 188
    with init_on_cpu():
        div_res = global_step / decay_steps
        if staircase:
189
            div_res = ops.floor(div_res)
190 191

        decayed_lr = learning_rate / (1 + decay_rate * div_res)
Q
Qiao Longfei 已提交
192

193
    return decayed_lr
194 195 196 197 198 199 200 201 202


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

Y
Yu Yang 已提交
203 204 205 206 207 208 209
    >>> 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
210 211 212 213 214 215 216 217 218 219 220
    Args:
        learning_rate: A scalar float32 value or a Variable. This
          will be the initial learning rate during training
        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
    """
Y
Yu Yang 已提交
221
    global_step = _decay_step_counter()
222

223 224
    with init_on_cpu():
        if cycle:
225 226
            div_res = ops.ceil(global_step / decay_steps)
            zero_var = tensor.fill_constant(
227
                shape=[1], dtype='float32', value=0.0)
228
            one_var = tensor.fill_constant(
229 230
                shape=[1], dtype='float32', value=1.0)

231
            with control_flow.Switch() as switch:
232
                with switch.case(global_step == zero_var):
233
                    tensor.assign(input=one_var, output=div_res)
234 235
            decay_steps = decay_steps * div_res
        else:
236
            decay_steps_var = tensor.fill_constant(
237
                shape=[1], dtype='float32', value=float(decay_steps))
238
            global_step = ops.elementwise_min(x=global_step, y=decay_steps_var)
239 240 241 242

        decayed_lr = (learning_rate - end_learning_rate) * \
                     ((1 - global_step / decay_steps) ** power) + end_learning_rate
    return decayed_lr
243 244


Y
Yu Yang 已提交
245
def piecewise_decay(boundaries, values):
246 247
    """Applies piecewise decay to the initial learning rate.

Y
Yu Yang 已提交
248 249 250 251 252 253 254 255 256
    >>> boundaries = [10000, 20000]
    >>> values = [1.0, 0.5, 0.1]
    >>>
    >>> if step < 10000:
    >>>     learning_rate = 1.0
    >>> elif 10000 <= step < 20000:
    >>>     learning_rate = 0.5
    >>> else:
    >>>     learning_rate = 0.1
257 258 259 260 261
    """

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

Y
Yu Yang 已提交
262
    global_step = _decay_step_counter()
263

264
    with init_on_cpu():
265
        lr = tensor.create_global_var(
266 267 268 269 270 271
            shape=[1],
            value=0.0,
            dtype='float32',
            persistable=True,
            name="learning_rate")

272
        with control_flow.Switch() as switch:
273
            for i in range(len(boundaries)):
274
                boundary_val = tensor.fill_constant(
275
                    shape=[1], dtype='float32', value=float(boundaries[i]))
276
                value_var = tensor.fill_constant(
277
                    shape=[1], dtype='float32', value=float(values[i]))
278
                with switch.case(global_step < boundary_val):
279 280
                    tensor.assign(value_var, lr)
            last_value_var = tensor.fill_constant(
281 282 283 284
                shape=[1],
                dtype='float32',
                value=float(values[len(values) - 1]))
            with switch.default():
285
                tensor.assign(last_value_var, lr)
286 287

    return lr