test_parallel_op.py 7.3 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.

Y
Yang Yang 已提交
15
import unittest
16

Y
Yang Yang 已提交
17
import paddle.v2.fluid as fluid
Y
Yang Yu 已提交
18 19 20 21
import numpy


class BaseParallelForTest(unittest.TestCase):
Y
Yang Yu 已提交
22 23 24 25 26 27 28 29 30
    def run_test(self, callback, feed, fetch):
        """
        Run the unittest for parallel.for
        Args:
            callback(callable): A callable function returns a generator. There 
                are two yields in the generator function. The first yield 
                returns the data layers, and the second yield returns the loss. 
                The modified data variables will be sent back during the first 
                yield.
31

Y
Yang Yu 已提交
32 33 34 35 36
            feed(dict): The executor feeding dictionary.
            fetch(list|basestr): The fetch name lists. 

        Returns:
            None
37

Y
Yang Yu 已提交
38 39 40 41 42
        Raises:
            AssertionError when the computation of cpu, parallel.for in cpu, 
                gpu, parallel.for in gpu are different.

        """
Y
Yang Yu 已提交
43
        cpu = fluid.CPUPlace()
Y
Yang Yu 已提交
44
        result_cpu = self._run_test_impl_(
Y
Yang Yu 已提交
45 46 47 48 49
            callback=callback,
            feed=feed,
            fetch=fetch,
            place=cpu,
            use_parallel=False)
Y
Yang Yu 已提交
50
        result_cpu_parallel = self._run_test_impl_(
Y
Yang Yu 已提交
51 52 53 54 55
            callback=callback,
            feed=feed,
            fetch=fetch,
            place=cpu,
            use_parallel=True)
56
        if fluid.core.is_compiled_with_cuda():
Y
Yang Yu 已提交
57
            gpu = fluid.CUDAPlace(0)
Y
Yang Yu 已提交
58
            result_gpu = self._run_test_impl_(
Y
Yang Yu 已提交
59 60 61 62 63
                callback=callback,
                feed=feed,
                fetch=fetch,
                place=gpu,
                use_parallel=False)
Y
Yang Yu 已提交
64
            result_gpu_parallel = self._run_test_impl_(
Y
Yang Yu 已提交
65 66 67 68 69
                callback=callback,
                feed=feed,
                fetch=fetch,
                place=gpu,
                use_parallel=True)
Y
Yang Yang 已提交
70 71 72 73 74 75 76
            result_gpu_nccl = self._run_test_impl_(
                callback=callback,
                feed=feed,
                fetch=fetch,
                place=gpu,
                use_parallel=True,
                use_nccl=True)
Y
Yang Yu 已提交
77
            self._assert_same_(fetch, result_cpu, result_cpu_parallel,
Y
Yang Yang 已提交
78
                               result_gpu, result_gpu_parallel, result_gpu_nccl)
Y
Yang Yu 已提交
79 80
        else:
            self._assert_same_(fetch, result_cpu, result_cpu_parallel)
Y
Yang Yu 已提交
81

Y
Yang Yang 已提交
82 83 84 85 86 87 88
    def _run_test_impl_(self,
                        callback,
                        feed,
                        fetch,
                        place,
                        use_parallel=False,
                        use_nccl=False):
Y
Yang Yu 已提交
89 90 91 92 93 94 95 96 97 98
        """
        Run a single test, returns the fetch values
        Args:
            place(Place): the computation place. 
            use_parallel(bool): Whether use parallel.for or not. 

        Returns:
            Fetched numpy arrays.

        """
Y
Yang Yu 已提交
99 100
        if isinstance(fetch, basestring):
            fetch = [fetch]
Y
Yang Yu 已提交
101 102 103 104 105 106 107 108 109 110 111
        main = fluid.Program()
        startup = fluid.Program()
        # Fix seed
        main.random_seed = 10
        startup.random_seed = 10

        with fluid.program_guard(main, startup):
            generator = callback()
            # Automatically insert parallel do if use_parallel = True
            if use_parallel:
                places = fluid.layers.get_places()
Y
Yang Yang 已提交
112
                pd = fluid.layers.ParallelDo(places, use_nccl=use_nccl)
Y
Yang Yu 已提交
113 114 115 116
                data = next(generator)

                if isinstance(data, fluid.Variable):
                    data = [data]
Y
Yang Yu 已提交
117

Y
Yang Yu 已提交
118 119 120 121
                with pd.do():
                    ins = map(pd.read_input, data)
                    if len(ins) == 1:
                        ins = ins[0]
Y
Yang Yu 已提交
122
                    loss = generator.send(ins)  # patch input
Y
Yang Yu 已提交
123 124 125 126 127
                    pd.write_output(loss)

                loss = pd()
            else:
                data = next(generator)
Y
Yang Yu 已提交
128 129
                loss = generator.send(data)
            self.assertIsNotNone(loss)
Y
Yang Yu 已提交
130 131 132 133 134 135 136
            avg_loss = fluid.layers.mean(x=loss)
            fluid.backward.append_backward(loss=avg_loss)

        exe = fluid.Executor(place)
        exe.run(startup)
        return exe.run(main, feed=feed, fetch_list=fetch)

Y
Yang Yu 已提交
137
    def _assert_same_(self, fetch, *args):
Y
Yang Yu 已提交
138 139 140 141 142 143 144 145 146 147 148 149 150 151
        """
        Assert the return values of `run_test` are same.
        Args:
            fetch: Fetch list. Used for print error message
            *args: The fetch result lists of each situations.

        Returns:
            None
            
        Raises:
            AssertionError

        """

Y
Yang Yu 已提交
152
        def _impl_(a, b, fetch_id, item_id):
Y
Yang Yang 已提交
153 154 155
            item_str = [
                'CPU', 'ParallelCPU', 'GPU', 'ParallelGPU', 'ParallelGPUNCCL'
            ]
156 157 158 159
            flag = numpy.allclose(a, b, rtol=0.1, atol=1e-3)
            self.assertTrue(flag,
                            "The {0} are different in {1}, {2} vs {3}".format(
                                fetch[fetch_id], item_str[item_id], a, b))
Y
Yang Yu 已提交
160 161 162 163 164 165

        for i, items in enumerate(zip(*args)):
            self.assertGreater(len(items), 0)
            for j in range(1, len(items)):
                _impl_(items[0], items[j], fetch_id=i, item_id=j)

Y
Yang Yu 已提交
166 167

class ParallelOpTest(BaseParallelForTest):
Y
Yu Yang 已提交
168 169 170 171 172 173 174
    @staticmethod
    def __network__():
        x = fluid.layers.data(shape=[784], dtype='float32', name='img')
        x = yield x
        hidden = fluid.layers.fc(input=x, size=200, param_attr='fc1.w')
        loss = fluid.layers.mean(x=hidden)
        yield loss
Y
Yang Yu 已提交
175

Y
Yang Yang 已提交
176
    def test_simple_fc(self):
Y
Yu Yang 已提交
177
        self.run_test(
Y
Yang Yang 已提交
178
            callback=self.__network__,
Y
Yang Yang 已提交
179 180 181
            feed={
                'img': numpy.random.random(size=(51, 784)).astype('float32')
            },
Y
Yang Yang 已提交
182
            fetch=['fc1.w@GRAD'])
Y
Yu Yang 已提交
183

Y
Yang Yang 已提交
184 185 186 187 188 189
    def test_fc_with_tiny_data(self):
        self.run_test(
            callback=self.__network__,
            feed={'img': numpy.random.random(size=(1, 784)).astype('float32')},
            fetch=['fc1.w@GRAD'])

Y
Yang Yang 已提交
190

Y
Yang Yang 已提交
191 192 193
class ParallelOpTestMultipleInput(BaseParallelForTest):
    @staticmethod
    def __network__():
Y
Yang Yu 已提交
194 195 196 197
        x = fluid.layers.data(
            shape=[784], dtype='float32', name='img1', stop_gradient=False)
        y = fluid.layers.data(
            shape=[784], dtype='float32', name='img2', stop_gradient=False)
Y
Yang Yang 已提交
198 199
        yield [x, y]
        x = x + y
Y
Yang Yang 已提交
200 201 202 203
        hidden1 = fluid.layers.fc(input=x, size=200, param_attr='fc1.w')
        hidden2 = fluid.layers.fc(input=hidden1, size=200, param_attr='fc2.w')
        hidden3 = fluid.layers.fc(input=hidden2, size=200, param_attr='fc3.w')
        loss = fluid.layers.mean(x=hidden3)
Y
Yang Yang 已提交
204 205 206 207 208 209 210 211 212
        yield loss

    def test_simple_fc(self):
        self.run_test(
            callback=self.__network__,
            feed={
                'img1': numpy.random.random(size=(51, 784)).astype('float32'),
                'img2': numpy.random.random(size=(51, 784)).astype('float32')
            },
Y
Yang Yang 已提交
213
            fetch=['fc1.w@GRAD', 'fc2.w@GRAD', 'fc3.w@GRAD'])
Y
Yang Yang 已提交
214 215


Y
Yang Yang 已提交
216 217
if __name__ == '__main__':
    unittest.main()