executor.py 11.8 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# 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
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13 14
# 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.

D
dzhwinter 已提交
15
import numpy as np
Y
Yang Yu 已提交
16
import contextlib
17
from framework import Program, default_main_program, Variable
18 19
from . import core

X
xuwei06 已提交
20 21 22
__all__ = [
    'Executor', 'global_scope', 'scope_guard', 'switch_scope', 'fetch_var'
]
Y
Yu Yang 已提交
23

Y
Yu Yang 已提交
24 25
g_scope = core.Scope()

Y
Yu Yang 已提交
26

Y
Yang Yu 已提交
27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44
def global_scope():
    return g_scope


def switch_scope(scope):
    global g_scope
    ex = g_scope
    g_scope = scope
    return ex


@contextlib.contextmanager
def scope_guard(scope):
    ex = switch_scope(scope)
    yield
    switch_scope(ex)


D
dzhwinter 已提交
45 46 47 48 49
def as_numpy(tensor):
    if isinstance(tensor, list):
        return [as_numpy(t) for t in tensor]
    assert isinstance(tensor, core.LoDTensor)
    lod = tensor.lod()
50 51 52 53 54 55 56
    if len(lod) > 0:
        raise RuntimeError(
            "Some of your featched tensors hold LoD information. \
            They can not be completely cast to Python ndarray. \
            Please set the parameter 'return_numpy' as 'False' to \
            return LoDTensor itself directly.")
    return np.array(tensor)
D
dzhwinter 已提交
57 58


59 60 61 62 63 64 65 66 67 68 69 70
def has_feed_operators(block, feed_targets, feed_holder_name):
    """ Check whether the block already has feed operators.

    Return false if the block does not have any feed operators.
    If some feed operators have been prepended to the block, check that
    the info contained in these feed operators matches the feed_targets
    and feed_holder_name. Raise exception when any mismatch is found.
    Return true when the block has feed operators with matching info.

    Args:
        block: a block instance (typically global block of a program)
        feed_targets: a dictionary of {feed_target_name: feed_target_data}
X
xuwei06 已提交
71 72
        feed_holder_name: the name of the variable that holds the data of
            all feed targets. The type of this feed_holder variable is
73 74 75
            FEED_MINIBATCH, which is essentially vector<LoDTensor>.

    Returns:
X
xuwei06 已提交
76
        A boolean value that indicates whether a block has feed operators
77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98
        that match the info contained in feed_targets and feed_holder_name.
    """

    feed_count = 0
    for op in block.ops:
        if op.desc.type() == 'feed':
            feed_count += 1
            assert op.desc.input('X')[0] == feed_holder_name
            feed_target_name = op.desc.output('Out')[0]
            if feed_target_name not in feed_targets:
                raise Exception("'feed_targets' does not have {} variable".
                                format(feed_target_name))
        else:
            break
    if feed_count > 0 and feed_count != len(feed_targets):
        raise Exception(
            "Feed operators in program desc do not match 'feed_targets'")
    return feed_count > 0


def has_fetch_operators(block, fetch_targets, fetch_holder_name):
    """ Check whether the block already has fetch operators.
X
xuwei06 已提交
99

100 101 102 103 104 105 106 107 108
    Return false if the block does not have any fetch operators.
    If some fetch operators have been appended to the block, check that
    the info contained in these fetch operators matches the fetch_targets
    and fetch_holder_name. Raise exception when any mismatch is found.
    Return true when the block has fetch operators with matching info.

    Args:
        block: a block instance (typically global block of a program)
        fetch_targets: a dictionary of {fetch_target_name: fetch_target_data}
X
xuwei06 已提交
109 110 111
        fetch_holder_name: the name of the variable that holds the data of
            all fetch targets. The type of this fetch_holder variable is
            FETCH_LIST, which is essentially vector<LoDTensor>.
112

X
xuwei06 已提交
113 114 115
    Return:
        A boolean value that indicates whether a block has fetch operators
        that match the info contained in fetch_targets and fetch_holder_name.
116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136
    """

    fetch_count = 0
    for op in block.ops:
        if op.desc.type() == 'fetch':
            fetch_count += 1
            assert op.desc.output('Out')[0] == fetch_holder_name
            fetch_target_name = op.desc.input('X')[0]
            if fetch_target_name not in [
                    var.desc.name() for var in fetch_targets
            ]:
                raise Exception("'fetch_targets' does not have {} variable".
                                format(fetch_target_name))
            idx = op.desc.attr('col')
            assert fetch_target_name == fetch_targets[idx].desc.name()
    if fetch_count > 0 and fetch_count != len(fetch_targets):
        raise Exception(
            "Fetch operators in program desc do not match 'fetch_targets'")
    return fetch_count > 0


X
xuwei06 已提交
137 138 139 140
def fetch_var(name, scope=None, return_numpy=True):
    """
    Fetch the value of the variable with the given name from the given scope
    Args:
141 142 143 144
        name(str): name of the variable. Typically, only persistable variables
            can be found in the scope used for running the program.
        scope(core.Scope|None): scope object. It should be the scope where
            you pass to Executor.run() when running your program.
X
xuwei06 已提交
145 146 147 148 149 150 151 152 153 154 155
            If None, global_scope() will be used.
        return_numpy(bool): whether convert the tensor to numpy.ndarray
    Returns:
       LodTensor|numpy.ndarray
    """
    assert isinstance(name, str)
    if scope is None:
        scope = global_scope()
    assert isinstance(scope, core.Scope)

    var = global_scope().find_var(name)
156 157 158 159
    assert var is not None, (
        "Cannot find " + name + " in scope. Perhaps you need to make the"
        " variable persistable by using var.persistable = True in your"
        " program.")
X
xuwei06 已提交
160 161 162 163 164 165
    tensor = var.get_tensor()
    if return_numpy:
        tensor = as_numpy(tensor)
    return tensor


Y
Yu Yang 已提交
166 167 168 169 170 171 172 173 174 175 176
class Executor(object):
    def __init__(self, places):
        if not isinstance(places, list) and not isinstance(places, tuple):
            places = [places]

        act_places = []
        for each in places:
            p = core.Place()
            p.set_place(each)
            act_places.append(p)

D
dzhwinter 已提交
177 178
        # TODO(dzhwinter) : only use the first place
        self.executor = core.Executor(act_places[0])
D
dzhwinter 已提交
179
        self.places = places
Q
qiaolongfei 已提交
180
        self.program_caches = dict()
D
dzhwinter 已提交
181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220

    def aslodtensor(self, data):
        def accumulate(data):
            if not isinstance(data, list):
                return 1
            return sum([accumulate(sub) for sub in data])

        def parselod(data):
            seq_lens = [accumulate(seq) for seq in data]
            cur_len = 0
            lod = [cur_len]
            for l in seq_lens:
                cur_len += l
                lod.append(cur_len)
            return lod

        assert len(self.places) != 0
        if not isinstance(data, list):
            # pure tensor case
            tensor = core.LoDTensor()
            tensor.set(data, self.places[0])
            return tensor
        else:
            raise RuntimeError("Current implementation lacks unittests")
            # lodtensor case
            lod = []
            if not isinstance(data[0], list):
                lod.append(parselod(data))
                flattened_data = np.concatenate(data, axis=0).astype("int64")
            else:
                while isinstance(data[0], list):
                    lod.append(parselod(seq))
                    flattened_data = [item for seq in data for item in seq]
                    data = flattened_data
                flattened_data = np.concatenate(data, axis=0).astype("int64")
            flattened_data = flattened_data.reshape([len(flattened_data), 1])
            tensor = core.LoDTensor()
            tensor.set(flattened_data, self.places[0])
            tensor.set_lod(lod)
            return tensor
Y
Yu Yang 已提交
221 222

    def run(self,
Y
Yu Yang 已提交
223
            program=None,
224 225
            feed=None,
            fetch_list=None,
Y
Yu Yang 已提交
226
            feed_var_name='feed',
Y
Yu Yang 已提交
227
            fetch_var_name='fetch',
D
dzhwinter 已提交
228
            scope=None,
229 230 231 232 233 234 235 236 237 238 239 240 241
            return_numpy=True,
            use_program_cache=False):
        """
        :param program: the program that need to run
        :param feed: feed variable list
        :param fetch_list: fetch variable list
        :param feed_var_name: feed_var_name default to 'feed'
        :param fetch_var_name: fetch_var_name default to 'fetch'
        :param scope: the scope used to run this program, you can switch it to different scope.
        :param return_numpy: convert the fetched tensor to numpy
        :param use_program_cache: set use_program_cache to true if program not changed compare to the last step.
        :return:
        """
242 243 244 245 246
        if feed is None:
            feed = {}
        if fetch_list is None:
            fetch_list = []

Y
Yu Yang 已提交
247
        if program is None:
Y
Yu Yang 已提交
248
            program = default_main_program()
Y
Yu Yang 已提交
249

Y
Yu Yang 已提交
250 251 252
        if not isinstance(program, Program):
            raise TypeError()

Y
Yu Yang 已提交
253
        if scope is None:
Y
Yang Yu 已提交
254
            scope = global_scope()
Y
Yu Yang 已提交
255

Q
qiaolongfei 已提交
256 257
        program_cache_key = str(feed.keys() + fetch_list)
        program_cache = self.program_caches.get(program_cache_key, None)
258

259
        if program_cache is None or not use_program_cache:
Q
qiaolongfei 已提交
260 261
            program_cache = program.clone()
            self.program_caches[program_cache_key] = program_cache
262

Q
qiaolongfei 已提交
263 264 265 266 267 268 269 270 271 272 273 274
            global_block = program_cache.global_block()

            if feed_var_name in global_block.vars:
                feed_var = global_block.var(feed_var_name)
            else:
                feed_var = global_block.create_var(
                    name=feed_var_name,
                    type=core.VarDesc.VarType.FEED_MINIBATCH,
                    persistable=True)

            if fetch_var_name in global_block.vars:
                fetch_var = global_block.var(fetch_var_name)
275
            else:
Q
qiaolongfei 已提交
276 277 278 279 280
                fetch_var = global_block.create_var(
                    name=fetch_var_name,
                    type=core.VarDesc.VarType.FETCH_LIST,
                    persistable=True)

281
            # prepend feed operators
Q
qiaolongfei 已提交
282 283 284 285 286 287 288 289 290
            if not has_feed_operators(global_block, feed, feed_var_name):
                for i, name in enumerate(feed):
                    out = global_block.var(name)
                    global_block.prepend_op(
                        type='feed',
                        inputs={'X': [feed_var]},
                        outputs={'Out': [out]},
                        attrs={'col': i})

291
            # append fetch_operators
Q
qiaolongfei 已提交
292 293 294 295 296 297 298 299 300 301 302
            if not has_fetch_operators(global_block, fetch_list,
                                       fetch_var_name):
                for i, var in enumerate(fetch_list):
                    assert isinstance(var, Variable) or isinstance(var, str), (
                        "Wrong type for fetch_list[%s]: %s" % (i, type(var)))
                    global_block.append_op(
                        type='fetch',
                        inputs={'X': [var]},
                        outputs={'Out': [fetch_var]},
                        attrs={'col': i})

Q
qiaolongfei 已提交
303 304 305 306 307 308 309 310 311 312 313 314
        # feed var to framework
        for op in program_cache.global_block().ops:
            if op.desc.type() == 'feed':
                feed_target_name = op.desc.output('Out')[0]
                cur_feed = feed[feed_target_name]
                if not isinstance(cur_feed, core.LoDTensor):
                    cur_feed = self.aslodtensor(cur_feed)
                idx = op.desc.attr('col')
                core.set_feed_variable(scope, cur_feed, feed_var_name, idx)
            else:
                break

Q
qiaolongfei 已提交
315
        self.executor.run(program_cache.desc, scope, 0, True, True)
D
dzhwinter 已提交
316
        outs = [
Y
Yu Yang 已提交
317
            core.get_fetch_variable(scope, fetch_var_name, i)
Y
Yu Yang 已提交
318 319
            for i in xrange(len(fetch_list))
        ]
D
dzhwinter 已提交
320 321 322
        if return_numpy:
            outs = as_numpy(outs)
        return outs