executor.py 13.0 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
    if len(lod) > 0:
D
dzhwinter 已提交
51
        raise RuntimeError("Some of your fetched tensors hold LoD information. \
52 53 54 55
            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
56 57


58 59 60 61 62 63 64 65 66 67 68 69
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
70 71
        feed_holder_name: the name of the variable that holds the data of
            all feed targets. The type of this feed_holder variable is
72 73 74
            FEED_MINIBATCH, which is essentially vector<LoDTensor>.

    Returns:
X
xuwei06 已提交
75
        A boolean value that indicates whether a block has feed operators
76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97
        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
98

99 100 101 102 103 104 105 106 107
    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
108 109 110
        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>.
111

X
xuwei06 已提交
112 113 114
    Return:
        A boolean value that indicates whether a block has fetch operators
        that match the info contained in fetch_targets and fetch_holder_name.
115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135
    """

    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
136 137 138 139
def fetch_var(name, scope=None, return_numpy=True):
    """
    Fetch the value of the variable with the given name from the given scope
    Args:
140 141 142 143
        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
144 145 146 147 148 149 150 151 152 153
            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)

Y
Yibing Liu 已提交
154
    var = scope.find_var(name)
155 156 157 158
    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
159 160 161 162 163 164
    tensor = var.get_tensor()
    if return_numpy:
        tensor = as_numpy(tensor)
    return tensor


Q
qiaolongfei 已提交
165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180
def get_program_cache_key(feed, fetch_list):
    feed_var_names = feed.keys()

    def to_name_str(var):
        if isinstance(var, Variable):
            return var.desc.name()
        elif isinstance(var, str):
            return var
        else:
            raise TypeError(str(var) + " should be Variable or str")

    fetch_var_names = map(to_name_str, fetch_list)

    return str(feed_var_names + fetch_var_names)


Y
Yu Yang 已提交
181
class Executor(object):
D
dzhwinter 已提交
182 183 184 185 186
    def __init__(self, place):
        self.place = place
        p = core.Place()
        p.set_place(place)
        self.executor = core.Executor(p)
Q
qiaolongfei 已提交
187
        self.program_caches = dict()
D
dzhwinter 已提交
188

D
dzhwinter 已提交
189 190 191 192 193 194 195 196 197 198 199
    def as_lodtensor(self, data):
        if isinstance(data, list):
            raise RuntimeError("Some of your feed data hold LoD information. \
                They can not be completely cast from a list of Python \
                ndarray to LoDTensor. Please convert data to LoDTensor \
                directly before feeding the data.\
                ")
        # single tensor case
        tensor = core.LoDTensor()
        tensor.set(data, self.place)
        return tensor
Y
Yu Yang 已提交
200

Q
Qiao Longfei 已提交
201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258
    def _get_program_cache(self, program_cache_key):
        return self.program_caches.get(program_cache_key, None)

    def _add_program_cache(self, program_cache_key, program):
        self.program_caches[program_cache_key] = program

    def _add_feed_fetch_ops(self, program, feed, fetch_list, feed_var_name,
                            fetch_var_name):
        tmp_program = program.clone()

        global_block = tmp_program.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)
        else:
            fetch_var = global_block.create_var(
                name=fetch_var_name,
                type=core.VarDesc.VarType.FETCH_LIST,
                persistable=True)

        # prepend feed operators
        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})

        # append fetch_operators
        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})

        return tmp_program

    def _feed_data(self, program, feed, feed_var_name, scope):
        # feed var to framework
        for op in program.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):
D
dzhwinter 已提交
259
                    cur_feed = self.as_lodtensor(cur_feed)
Q
Qiao Longfei 已提交
260 261 262 263 264 265 266 267 268 269 270 271
                idx = op.desc.attr('col')
                core.set_feed_variable(scope, cur_feed, feed_var_name, idx)
            else:
                break

    def _fetch_data(self, fetch_list, fetch_var_name, scope):
        outs = [
            core.get_fetch_variable(scope, fetch_var_name, i)
            for i in xrange(len(fetch_list))
        ]
        return outs

Y
Yu Yang 已提交
272
    def run(self,
Y
Yu Yang 已提交
273
            program=None,
274 275
            feed=None,
            fetch_list=None,
Y
Yu Yang 已提交
276
            feed_var_name='feed',
Y
Yu Yang 已提交
277
            fetch_var_name='fetch',
D
dzhwinter 已提交
278
            scope=None,
279 280
            return_numpy=True,
            use_program_cache=False):
Q
qiaolongfei 已提交
281 282 283 284
        """ Run program by this Executor. Feed data by feed map, fetch result by fetch_list.

        Python executor takes a program, add feed operators and fetch operators to this program according
        to feed map and fetch_list. Feed map provides input data for the program. fetch_list provides
Q
qiaolongfei 已提交
285
        the variables(or names) that user want to get after program run. Note: the executor will run all
Q
qiaolongfei 已提交
286 287 288 289
        operators in the program but not only the operators dependent by the fetch_list

        :param program: the program that need to run, if not provied, then default_main_program will be used.
        :param feed: feed variable map, e.g. {"image": ImageData, "label": LableData}
Q
qiaolongfei 已提交
290 291
        :param fetch_list: a list of variable or variable names that user want to get, run will return them according
        to this list.
Q
qiaolongfei 已提交
292 293 294 295
        :param feed_var_name: the name for the input variable of feed Operator.
        :param fetch_var_name: the name for the output variable of feed Operator.
        :param scope: the scope used to run this program, you can switch it to different scope. default is global_scope
        :param return_numpy: if convert the fetched tensor to numpy
296
        :param use_program_cache: set use_program_cache to true if program not changed compare to the last step.
Q
qiaolongfei 已提交
297
        :return: result according to fetch_list.
298
        """
299 300
        if feed is None:
            feed = {}
Q
qiaolongfei 已提交
301
        if not isinstance(feed, dict):
D
dzhwinter 已提交
302 303 304
            raise TypeError(
                "feed requires dict as its Parameter. But you passed in %s" %
                (type(feed)))
305 306
        if fetch_list is None:
            fetch_list = []
Y
Yu Yang 已提交
307
        if program is None:
Y
Yu Yang 已提交
308
            program = default_main_program()
Y
Yu Yang 已提交
309

Y
Yu Yang 已提交
310
        if not isinstance(program, Program):
D
dzhwinter 已提交
311 312 313
            raise TypeError(
                "Executor requires Program as its Parameter. But you passed in %s"
                % (type(program)))
Y
Yu Yang 已提交
314

Y
Yu Yang 已提交
315
        if scope is None:
Y
Yang Yu 已提交
316
            scope = global_scope()
Y
Yu Yang 已提交
317

Q
Qiao Longfei 已提交
318
        cache_key = get_program_cache_key(feed, fetch_list)
319
        if use_program_cache:
Q
Qiao Longfei 已提交
320 321 322 323 324 325 326 327 328 329
            cached_program = self._get_program_cache(cache_key)
            if cached_program is None:
                cached_program = self._add_feed_fetch_ops(
                    program=program,
                    feed=feed,
                    fetch_list=fetch_list,
                    feed_var_name=feed_var_name,
                    fetch_var_name=fetch_var_name)
                self._add_program_cache(cache_key, cached_program)
            program = cached_program
330
        else:
Q
Qiao Longfei 已提交
331 332 333 334 335 336 337 338 339 340 341
            self.program_caches.pop(cache_key, None)
            program = self._add_feed_fetch_ops(
                program=program,
                feed=feed,
                fetch_list=fetch_list,
                feed_var_name=feed_var_name,
                fetch_var_name=fetch_var_name)

        self._feed_data(program, feed, feed_var_name, scope)
        self.executor.run(program.desc, scope, 0, True, True)
        outs = self._fetch_data(fetch_list, fetch_var_name, scope)
D
dzhwinter 已提交
342 343 344
        if return_numpy:
            outs = as_numpy(outs)
        return outs