trainer.py 6.1 KB
Newer Older
Y
Yu Yang 已提交
1
import collections
Y
Yu Yang 已提交
2

Y
Yu Yang 已提交
3
import py_paddle.swig_paddle as api
Q
qiaolongfei 已提交
4
from paddle.proto.ModelConfig_pb2 import ModelConfig
Y
Yu Yang 已提交
5 6
from py_paddle import DataProviderConverter

Q
qiaolongfei 已提交
7 8
from . import event as v2_event
from . import layer as v2_layer
Y
Yu Yang 已提交
9 10 11
from . import optimizer as v2_optimizer
from . import parameters as v2_parameters

Y
Yu Yang 已提交
12
__all__ = ['ITrainer', 'SGD']
Y
Yu Yang 已提交
13 14 15


def default_event_handler(event):
Y
Yu Yang 已提交
16 17 18 19 20 21 22
    """
    Default event handler. It will print some log and save mode.

    TODO(yuyang18): Complete it!
    :param event:
    :return:
    """
Y
Yu Yang 已提交
23 24 25 26
    pass


class ITrainer(object):
Y
Yu Yang 已提交
27 28 29 30
    """
    The interface of Trainer. The only exposed method is `train`.
    """

Y
Yu Yang 已提交
31 32 33 34 35 36
    def train(self,
              train_data_reader,
              topology,
              parameters,
              test_data_reader=None,
              event_handler=None):
Y
Yu Yang 已提交
37 38 39 40 41 42 43 44 45 46 47
        """
        train method.

        :param train_data_reader:
        :param topology:
        :param parameters:
        :param test_data_reader:
        :param event_handler:
        :return:
        """

Y
Yu Yang 已提交
48 49 50
        raise NotImplementedError()


Y
Yu Yang 已提交
51
class SGD(ITrainer):
Y
Yu Yang 已提交
52
    def __init__(self, update_equation):
Y
Yu Yang 已提交
53 54 55
        """
        Simple SGD Trainer.

Y
Yu Yang 已提交
56 57
        :param update_equation: The optimizer object.
        :type update_equation: v2_optimizer.Optimizer
Y
Yu Yang 已提交
58
        """
Y
Yu Yang 已提交
59 60 61
        if not isinstance(update_equation, v2_optimizer.Optimizer):
            raise ValueError("update equation parameter must be "
                             "paddle.v2.optimizer.Optimizer")
Y
Yu Yang 已提交
62 63 64 65 66 67 68 69 70 71 72
        self.__optimizer__ = update_equation

    def train(self,
              train_data_reader,
              topology,
              parameters,
              num_passes=1,
              test_data_reader=None,
              event_handler=None,
              batch_size=32,
              data_types=None):
Y
Yu Yang 已提交
73 74 75 76
        """
        Training method. Will train num_passes of input data.

        :param train_data_reader:
Q
qiaolongfei 已提交
77
        :param topology: Network Topology, use one or more Layers to represent it.
Y
Yu Yang 已提交
78 79 80 81 82 83 84 85 86 87
        :param parameters: The parameter pools.
        :param num_passes: The total train passes.
        :param test_data_reader:
        :param event_handler: Event handler. A method will be invoked when event
                              occurred.
        :type event_handler: (BaseEvent) => None
        :param batch_size: Not important, will be removed after data refactor.
        :param data_types: Not important, will be removed after data refactor.
        :return:
        """
Y
Yu Yang 已提交
88 89 90
        if event_handler is None:
            event_handler = default_event_handler

Q
qiaolongfei 已提交
91 92
        topology = v2_layer.parse_network(topology)

Y
Yu Yang 已提交
93 94 95 96 97
        __check_train_args__(**locals())

        gm = api.GradientMachine.createFromConfigProto(
            topology, api.CREATE_MODE_NORMAL, self.__optimizer__.enable_types())
        assert isinstance(gm, api.GradientMachine)
Y
Yu Yang 已提交
98
        parameters.append_gradient_machine(gm)
Y
Yu Yang 已提交
99 100 101 102

        updater = self.__optimizer__.create_local_updater()
        updater.init(gm)

Y
Yu Yang 已提交
103 104 105
        gm.start()
        out_args = api.Arguments.createArguments(0)

Y
Yu Yang 已提交
106 107 108 109 110 111 112 113 114 115 116
        data_types_lists = []
        for each in topology.input_layer_names:
            if each not in data_types:
                raise ValueError()
            data_types_lists.append(data_types[each])

        converter = DataProviderConverter(input_types=data_types_lists)

        for pass_id in xrange(num_passes):
            updater.startPass()
            for batch_id, data_batch in enumerate(
Y
Yu Yang 已提交
117 118
                    __data_reader_to_batch__(train_data_reader, batch_size,
                                             topology)):
Y
Yu Yang 已提交
119 120 121 122 123 124 125 126 127 128
                pass_type = updater.startBatch(len(data_batch))
                gm.forwardBackward(converter(data_batch), out_args, pass_type)
                for each_param in gm.getParameters():
                    updater.update(each_param)
                # Get cost. We use numpy to calculate total cost for this batch.
                cost_vec = out_args.getSlotValue(0)
                cost_vec = cost_vec.copyToNumpyMat()
                cost = cost_vec.sum() / len(data_batch)
                updater.finishBatch(cost)
                event_handler(
Y
Yu Yang 已提交
129
                    v2_event.EndIteration(
Y
Yu Yang 已提交
130
                        pass_id=pass_id, batch_id=batch_id, cost=cost))
Y
Yu Yang 已提交
131 132 133 134 135

            updater.finishPass()
        gm.finish()


Y
Yu Yang 已提交
136 137 138 139 140 141 142 143 144 145 146 147 148 149 150
def __data_reader_to_batch__(reader, batch_size, topology):
    """
    This function is not important, and will be removed when data refactored.
    """

    def input_reorder(func):
        for item in func():
            retv = []
            for __layer_name__ in topology.input_layer_names:
                retv.append(item[__layer_name__])
            yield retv

    return __generator_to_batch__(input_reorder(reader), batch_size=batch_size)


Y
Yu Yang 已提交
151
def __generator_to_batch__(generator, batch_size):
Y
Yu Yang 已提交
152 153 154
    """
    This function is not important, and will be removed when data refactored.
    """
Y
Yu Yang 已提交
155 156 157 158 159 160 161 162 163 164 165 166
    ret_val = list()
    for each_item in generator:
        ret_val.append(each_item)
        if len(ret_val) == batch_size:
            yield ret_val
            ret_val = list()
    if len(ret_val) != 0:
        yield ret_val


def __check_train_args__(train_data_reader, topology, parameters,
                         test_data_reader, event_handler, **kwargs):
Y
Yu Yang 已提交
167 168 169
    """
    Check train function's argument types
    """
Y
Yu Yang 已提交
170 171 172 173 174 175 176 177 178 179 180 181 182 183
    if not callable(train_data_reader) or not isinstance(train_data_reader(),
                                                         collections.Iterator):
        raise ValueError('train_data_reader should be a function, '
                         'which can return a iterator')

    if test_data_reader is not None:
        if not callable(test_data_reader) or not isinstance(
                test_data_reader(), collections.Iterator):
            raise ValueError('test_data_reader should be a function, which can '
                             'return a iterator')

    if not isinstance(topology, ModelConfig):
        raise ValueError('topology should be a model config')

Y
Yu Yang 已提交
184
    if not isinstance(parameters, v2_parameters.Parameters):
Y
Yu Yang 已提交
185 186 187 188
        raise ValueError('parameters should be a parameter pool')

    if not callable(event_handler):
        raise ValueError('event handler should be a function')