optimizer.py 9.5 KB
Newer Older
Q
qiaolongfei 已提交
1
import py_paddle.swig_paddle as swig_api
Y
Yu Yang 已提交
2

Q
qiaolongfei 已提交
3
import paddle.trainer_config_helpers.config_parser_utils as config_parser_utils
Y
Yu Yang 已提交
4 5 6 7
import paddle.trainer_config_helpers.optimizers as v1_optimizers
"""
Optimizers(update equation) for SGD method.

D
dzhwinter 已提交
8 9
TODO(zhihong) : create new optimizer with proto config, add new optimizer here

Y
Yu Yang 已提交
10 11
TODO(yuyang18): Complete comments.
"""
Q
qiaolongfei 已提交
12

L
Luo Tao 已提交
13 14 15 16
__all__ = [
    'Momentum', 'Adam', 'Adamax', 'AdaGrad', 'DecayedAdaGrad', 'AdaDelta',
    'RMSProp', 'ModelAverage', 'L2Regularization'
]
Q
qiaolongfei 已提交
17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42


class Optimizer(object):
    def __init__(self, **kwargs):
        if 'batch_size' in kwargs:
            del kwargs['batch_size']  # not important for python library.

        def __impl__():
            v1_optimizers.settings(batch_size=1, **kwargs)

        self.__opt_conf_proto__ = config_parser_utils.parse_optimizer_config(
            __impl__)
        self.__opt_conf__ = swig_api.OptimizationConfig.createFromProto(
            self.__opt_conf_proto__)

    def enable_types(self):
        """
        get enable_types for each optimizer.
        enable_types = [value, gradient, momentum, etc]
        For each optimizer(SGD, Adam), GradientMachine should enable different
        buffers.
        """
        tmp = swig_api.ParameterOptimizer.create(self.__opt_conf__)
        assert isinstance(tmp, swig_api.ParameterOptimizer)
        return tmp.getParameterTypes()

Q
qiaolongfei 已提交
43
    def __create_local_updater__(self):
Q
qiaolongfei 已提交
44 45
        return swig_api.ParameterUpdater.createLocalUpdater(self.__opt_conf__)

Q
qiaolongfei 已提交
46
    def __create_remote_updater__(self, pass_num, use_sparse_updater):
Q
qiaolongfei 已提交
47 48
        return swig_api.ParameterUpdater.createRemoteUpdater(
            self.__opt_conf__, pass_num, use_sparse_updater)
Q
qiaolongfei 已提交
49

50 51 52 53 54 55
    def __create_new_remote_updater__(self, pserver_spec):
        return swig_api.ParameterUpdater.createNewRemoteUpdater(
            self.__opt_conf__, pserver_spec)

    def create_updater(self, is_local, num_passes, use_sparse_updater,
                       pserver_spec):
Q
qiaolongfei 已提交
56 57 58 59 60 61 62 63 64 65 66 67 68
        """
        create proper parameter_updater by configuration.
        :param is_local: create local or remote parameter updater
        :param num_passes: remote parameter updater will use this to config
        parameter server.
        :param use_sparse_updater: when use remote updater, if some parameter is
        sparse, updater should do some extra thing:

        ..  code-block:: python

            if use_sparse_remote_updater:
                        gradient_machine.prefetch(in_args)
                        parameter_updater.getParametersRemote()
W
wuyi05 已提交
69 70

        :param pserver_spec: pserver location, eg: localhost:3000
Q
qiaolongfei 已提交
71 72
        :return: parameter_updater
        """
Q
qiaolongfei 已提交
73
        if is_local:
Q
qiaolongfei 已提交
74
            parameter_updater = self.__create_local_updater__()
Q
qiaolongfei 已提交
75
        else:
76 77 78 79 80 81
            if pserver_spec is None:
                parameter_updater = self.__create_remote_updater__(
                    num_passes, use_sparse_updater)
            else:
                parameter_updater = self.__create_new_remote_updater__(
                    pserver_spec)
Q
qiaolongfei 已提交
82
        return parameter_updater
Q
qiaolongfei 已提交
83 84


L
Luo Tao 已提交
85
class Momentum(Optimizer):
Q
qijun 已提交
86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114
    """
    SGD Optimizer.

    SGD is an optimization method, trying to find a neural network that
    minimize the "cost/error" of it by iteration. In paddle's implementation
    SGD Optimizer is synchronized, which means all gradients will be wait to
    calculate and reduced into one gradient, then do optimize operation.

    The neural network consider the learning problem of minimizing an objective
    function, that has the form of a sum

    ..  math::

        Q(w) = \\sum_{i}^{n} Q_i(w)

    The value of function Q sometimes is the cost of neural network (Mean
    Square Error between prediction and label for example). The function Q is
    parametrised by w, the weight/bias of neural network. And weights is what to
    be learned. The i is the i-th observation in (trainning) data.

    So, the SGD method will optimize the weight by

    ..  math::

        w = w - \\eta \\nabla Q(w) = w - \\eta \\sum_{i}^{n} \\nabla Q_i(w)

    where :math:`\\eta` is learning rate. And :math:`n` is batch size.
    """

L
Luo Tao 已提交
115 116
    def __init__(self, momentum=None, sparse=False, **kwargs):
        learning_method = v1_optimizers.MomentumOptimizer(
Y
Yu Yang 已提交
117
            momentum=momentum, sparse=sparse)
L
Luo Tao 已提交
118 119 120 121
        super(Momentum, self).__init__(
            learning_method=learning_method, **kwargs)


Q
qiaolongfei 已提交
122
class Adam(Optimizer):
Q
qijun 已提交
123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142
    """
    Adam optimizer.
    The details of please refer `Adam: A Method for Stochastic Optimization
    <https://arxiv.org/abs/1412.6980>`_

    ..  math::

        m(w, t) & = \\beta_1 m(w, t-1) + (1 - \\beta_1) \\nabla Q_i(w) \\\\
        v(w, t) & = \\beta_2 v(w, t-1) + (1 - \\beta_2)(\\nabla Q_i(w)) ^2 \\\\
        w & = w - \\frac{\\eta}{\\sqrt{v(w,t) + \\epsilon}}

    :param beta1: the :math:`\\beta_1` in equation.
    :type beta1: float
    :param beta2: the :math:`\\beta_2` in equation.
    :type beta2: float
    :param epsilon: the :math:`\\epsilon` in equation. It is used to prevent
                        divided by zero.
    :type epsilon: float
    """

Q
qiaolongfei 已提交
143 144 145 146 147 148 149
    def __init__(self, beta1=0.9, beta2=0.999, epsilon=1e-8, **kwargs):
        learning_method = v1_optimizers.AdamOptimizer(
            beta1=beta1, beta2=beta2, epsilon=epsilon)
        super(Adam, self).__init__(learning_method=learning_method, **kwargs)


class Adamax(Optimizer):
Q
qijun 已提交
150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167
    """
    Adamax optimizer.

    The details of please refer this `Adam: A Method for Stochastic Optimization
    <https://arxiv.org/abs/1412.6980>`_

    ..  math::

        m_t & = \\beta_1 * m_{t-1} + (1-\\beta_1)* \\nabla Q_i(w) \\\\
        u_t & = max(\\beta_2*u_{t-1}, abs(\\nabla Q_i(w))) \\\\
        w_t & = w_{t-1} - (\\eta/(1-\\beta_1^t))*m_t/u_t

    :param beta1: the :math:`\\beta_1` in the equation.
    :type beta1: float
    :param beta2: the :math:`\\beta_2` in the equation.
    :type beta2: float
    """

Q
qiaolongfei 已提交
168 169 170 171 172 173
    def __init__(self, beta1=0.9, beta2=0.999, **kwargs):
        learning_method = v1_optimizers.AdamaxOptimizer(
            beta1=beta1, beta2=beta2)
        super(Adamax, self).__init__(learning_method=learning_method, **kwargs)


L
Luo Tao 已提交
174
class AdaGrad(Optimizer):
Q
qijun 已提交
175 176 177 178 179 180 181 182 183 184 185 186 187
    """
    Adagrad(for ADAptive GRAdient algorithm) optimizer.

    For details please refer this `Adaptive Subgradient Methods for
    Online Learning and Stochastic Optimization
    <http://www.magicbroom.info/Papers/DuchiHaSi10.pdf>`_.

    ..  math::

        G &= \\sum_{\\tau=1}^{t} g_{\\tau} g_{\\tau}^T \\\\
        w & = w - \\eta diag(G)^{-\\frac{1}{2}} \\circ g
    """

L
Luo Tao 已提交
188 189 190 191 192 193
    def __init__(self, **kwargs):
        learning_method = v1_optimizers.AdaGradOptimizer()
        super(AdaGrad, self).__init__(learning_method=learning_method, **kwargs)


class DecayedAdaGrad(Optimizer):
Q
qijun 已提交
194 195 196 197 198 199 200 201 202 203 204 205 206 207 208
    """
    AdaGrad method with decayed sum gradients. The equations of this method
    show as follow.

    ..  math::

        E(g_t^2) &= \\rho * E(g_{t-1}^2) + (1-\\rho) * g^2 \\\\
        learning\\_rate &= 1/sqrt( ( E(g_t^2) + \\epsilon )

    :param rho: The :math:`\\rho` parameter in that equation
    :type rho: float
    :param epsilon: The :math:`\\epsilon` parameter in that equation.
    :type epsilon: float
    """

L
Luo Tao 已提交
209 210 211 212 213 214 215 216
    def __init__(self, rho=0.95, epsilon=1e-06, **kwargs):
        learning_method = v1_optimizers.DecayedAdaGradOptimizer(
            rho=rho, epsilon=epsilon)
        super(DecayedAdaGrad, self).__init__(
            learning_method=learning_method, **kwargs)


class AdaDelta(Optimizer):
Q
qijun 已提交
217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233
    """
    AdaDelta method. The details of adadelta please refer to this
    `ADADELTA: AN ADAPTIVE LEARNING RATE METHOD
    <http://www.matthewzeiler.com/pubs/googleTR2012/googleTR2012.pdf>`_.

    ..  math::

        E(g_t^2) &= \\rho * E(g_{t-1}^2) + (1-\\rho) * g^2 \\\\
        learning\\_rate &= sqrt( ( E(dx_{t-1}^2) + \\epsilon ) / ( \\
                          E(g_t^2) + \\epsilon ) ) \\\\
        E(dx_t^2) &= \\rho * E(dx_{t-1}^2) + (1-\\rho) * (-g*learning\\_rate)^2

    :param rho: :math:`\\rho` in equation
    :type rho: float
    :param epsilon: :math:`\\rho` in equation
    :type epsilon: float
    """
Q
qijun 已提交
234

L
Luo Tao 已提交
235 236 237 238 239 240 241 242
    def __init__(self, rho=0.95, epsilon=1e-06, **kwargs):
        learning_method = v1_optimizers.AdaDeltaOptimizer(
            rho=rho, epsilon=epsilon)
        super(AdaDelta, self).__init__(
            learning_method=learning_method, **kwargs)


class RMSProp(Optimizer):
Q
qijun 已提交
243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260
    """
    RMSProp(for Root Mean Square Propagation) optimizer. For details please
    refer this `slide <http://www.cs.toronto.edu/~tijmen/csc321/slides/
    lecture_slides_lec6.pdf>`_.

    The equations of this method as follows:

    ..  math::

        v(w, t) & = \\rho v(w, t-1) + (1 - \\rho)(\\nabla Q_{i}(w))^2 \\\\
        w & = w - \\frac{\\eta} {\\sqrt{v(w,t) + \\epsilon}} \\nabla Q_{i}(w)

    :param rho: the :math:`\\rho` in the equation. The forgetting factor.
    :type rho: float
    :param epsilon: the :math:`\\epsilon` in the equation.
    :type epsilon: float
    """

L
Luo Tao 已提交
261 262 263 264 265 266 267 268 269
    def __init__(self, rho=0.95, epsilon=1e-6, **kwargs):
        learning_method = v1_optimizers.RMSPropOptimizer(
            rho=rho, epsilon=epsilon)
        super(RMSProp, self).__init__(learning_method=learning_method, **kwargs)


ModelAverage = v1_optimizers.ModelAverage
L2Regularization = v1_optimizers.L2Regularization

Q
qiaolongfei 已提交
270 271
if __name__ == '__main__':
    swig_api.initPaddle('--use_gpu=false')
L
Luo Tao 已提交
272 273 274 275 276 277 278 279
    for opt in [
            Momentum(), Adam(), Adamax(), AdaGrad(), DecayedAdaGrad(),
            AdaDelta(), RMSProp(), Adam(
                model_average=ModelAverage(average_window=0.5),
                regularization=L2Regularization(rate=0.5),
                gradient_clipping_threshold=25)
    ]:
        print opt, opt.enable_types()