test_imperative_ptb_rnn.py 15.1 KB
Newer Older
J
JiabinYang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# 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
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# 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.

from __future__ import print_function

import unittest
import paddle.fluid as fluid
L
lujun 已提交
19
from paddle.fluid.dygraph.nn import Embedding
J
JiabinYang 已提交
20 21
import paddle.fluid.framework as framework
from paddle.fluid.optimizer import SGDOptimizer
L
lujun 已提交
22
from paddle.fluid.dygraph.base import to_variable
23
from test_imperative_base import new_program_scope
J
JiabinYang 已提交
24
import numpy as np
25
import six
J
JiabinYang 已提交
26 27 28
from paddle.fluid.backward import append_backward


L
lujun 已提交
29
class SimpleLSTMRNN(fluid.dygraph.Layer):
J
JiabinYang 已提交
30
    def __init__(self,
X
Xin Pan 已提交
31
                 name_scope,
J
JiabinYang 已提交
32 33 34 35 36
                 hidden_size,
                 num_steps,
                 num_layers=2,
                 init_scale=0.1,
                 dropout=None):
X
Xin Pan 已提交
37
        super(SimpleLSTMRNN, self).__init__(name_scope)
J
JiabinYang 已提交
38 39 40 41
        self._hidden_size = hidden_size
        self._num_layers = num_layers
        self._init_scale = init_scale
        self._dropout = dropout
42 43
        self._input = None
        self._num_steps = num_steps
44 45
        self.cell_array = []
        self.hidden_array = []
J
JiabinYang 已提交
46 47 48 49 50 51 52 53

    def _build_once(self, input_embedding, init_hidden=None, init_cell=None):
        self.weight_1_arr = []
        self.weight_2_arr = []
        self.bias_arr = []
        self.mask_array = []

        for i in range(self._num_layers):
54
            weight_1 = self.create_parameter(
55 56 57
                attr=fluid.ParamAttr(
                    initializer=fluid.initializer.UniformInitializer(
                        low=-self._init_scale, high=self._init_scale)),
J
JiabinYang 已提交
58 59 60 61
                shape=[self._hidden_size * 2, self._hidden_size * 4],
                dtype="float32",
                default_initializer=fluid.initializer.UniformInitializer(
                    low=-self._init_scale, high=self._init_scale))
62
            self.weight_1_arr.append(self.add_parameter('w_%d' % i, weight_1))
63
            bias_1 = self.create_parameter(
64 65 66 67
                attr=fluid.ParamAttr(
                    initializer=fluid.initializer.UniformInitializer(
                        low=-self._init_scale, high=self._init_scale)),
                shape=[self._hidden_size * 4],
J
JiabinYang 已提交
68 69
                dtype="float32",
                default_initializer=fluid.initializer.Constant(0.0))
70
            self.bias_arr.append(self.add_parameter('b_%d' % i, bias_1))
J
JiabinYang 已提交
71

72 73 74 75 76
    def forward(self, input_embedding, init_hidden=None, init_cell=None):
        self.cell_array = []
        self.hidden_array = []

        for i in range(self._num_layers):
J
JiabinYang 已提交
77 78 79 80 81 82 83 84 85 86 87 88
            pre_hidden = fluid.layers.slice(
                init_hidden, axes=[0], starts=[i], ends=[i + 1])
            pre_cell = fluid.layers.slice(
                init_cell, axes=[0], starts=[i], ends=[i + 1])
            pre_hidden = fluid.layers.reshape(
                pre_hidden, shape=[-1, self._hidden_size])
            pre_cell = fluid.layers.reshape(
                pre_cell, shape=[-1, self._hidden_size])
            self.hidden_array.append(pre_hidden)
            self.cell_array.append(pre_cell)

        res = []
89 90
        for index in range(self._num_steps):
            self._input = fluid.layers.slice(
J
JiabinYang 已提交
91
                input_embedding, axes=[1], starts=[index], ends=[index + 1])
92 93
            self._input = fluid.layers.reshape(
                self._input, shape=[-1, self._hidden_size])
J
JiabinYang 已提交
94 95 96 97 98 99
            for k in range(self._num_layers):
                pre_hidden = self.hidden_array[k]
                pre_cell = self.cell_array[k]
                weight_1 = self.weight_1_arr[k]
                bias = self.bias_arr[k]

100
                nn = fluid.layers.concat([self._input, pre_hidden], 1)
J
JiabinYang 已提交
101 102 103
                gate_input = fluid.layers.matmul(x=nn, y=weight_1)

                gate_input = fluid.layers.elementwise_add(gate_input, bias)
104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131
                i, j, f, o = fluid.layers.split(
                    gate_input, num_or_sections=4, dim=-1)
                c = pre_cell * fluid.layers.sigmoid(f) + fluid.layers.sigmoid(
                    i) * fluid.layers.tanh(j)
                m = fluid.layers.tanh(c) * fluid.layers.sigmoid(o)
                self.hidden_array[k] = m
                self.cell_array[k] = c
                self._input = m

                if self._dropout is not None and self._dropout > 0.0:
                    self._input = fluid.layers.dropout(
                        self._input,
                        dropout_prob=self._dropout,
                        dropout_implementation='upscale_in_train')
            res.append(
                fluid.layers.reshape(
                    self._input, shape=[1, -1, self._hidden_size]))
        real_res = fluid.layers.concat(res, 0)
        real_res = fluid.layers.transpose(x=real_res, perm=[1, 0, 2])
        last_hidden = fluid.layers.concat(self.hidden_array, 1)
        last_hidden = fluid.layers.reshape(
            last_hidden, shape=[-1, self._num_layers, self._hidden_size])
        last_hidden = fluid.layers.transpose(x=last_hidden, perm=[1, 0, 2])
        last_cell = fluid.layers.concat(self.cell_array, 1)
        last_cell = fluid.layers.reshape(
            last_cell, shape=[-1, self._num_layers, self._hidden_size])
        last_cell = fluid.layers.transpose(x=last_cell, perm=[1, 0, 2])
        return real_res, last_hidden, last_cell
J
JiabinYang 已提交
132 133


L
lujun 已提交
134
class PtbModel(fluid.dygraph.Layer):
J
JiabinYang 已提交
135
    def __init__(self,
X
Xin Pan 已提交
136
                 name_scope,
J
JiabinYang 已提交
137 138 139 140 141 142
                 hidden_size,
                 vocab_size,
                 num_layers=2,
                 num_steps=20,
                 init_scale=0.1,
                 dropout=None):
X
Xin Pan 已提交
143
        super(PtbModel, self).__init__(name_scope)
J
JiabinYang 已提交
144 145 146 147 148 149 150
        self.hidden_size = hidden_size
        self.vocab_size = vocab_size
        self.init_scale = init_scale
        self.num_layers = num_layers
        self.num_steps = num_steps
        self.dropout = dropout
        self.simple_lstm_rnn = SimpleLSTMRNN(
X
Xin Pan 已提交
151
            self.full_name(),
J
JiabinYang 已提交
152 153 154 155 156
            hidden_size,
            num_steps,
            num_layers=num_layers,
            init_scale=init_scale,
            dropout=dropout)
157
        self.embedding = Embedding(
X
Xin Pan 已提交
158
            self.full_name(),
J
JiabinYang 已提交
159 160 161 162 163 164 165
            size=[vocab_size, hidden_size],
            dtype='float32',
            is_sparse=False,
            param_attr=fluid.ParamAttr(
                name='embedding_para',
                initializer=fluid.initializer.UniformInitializer(
                    low=-init_scale, high=init_scale)))
166
        self.softmax_weight = self.create_parameter(
167 168
            attr=fluid.ParamAttr(),
            shape=[self.hidden_size, self.vocab_size],
J
JiabinYang 已提交
169 170 171
            dtype="float32",
            default_initializer=fluid.initializer.UniformInitializer(
                low=-self.init_scale, high=self.init_scale))
172
        self.softmax_bias = self.create_parameter(
173 174
            attr=fluid.ParamAttr(),
            shape=[self.vocab_size],
J
JiabinYang 已提交
175 176 177 178
            dtype="float32",
            default_initializer=fluid.initializer.UniformInitializer(
                low=-self.init_scale, high=self.init_scale))

179 180 181
    def _build_once(self, input, label, init_hidden, init_cell):
        pass

J
JiabinYang 已提交
182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200
    def forward(self, input, label, init_hidden, init_cell):
        init_h = fluid.layers.reshape(
            init_hidden, shape=[self.num_layers, -1, self.hidden_size])

        init_c = fluid.layers.reshape(
            init_cell, shape=[self.num_layers, -1, self.hidden_size])

        x_emb = self.embedding(input)
        x_emb = fluid.layers.reshape(
            x_emb, shape=[-1, self.num_steps, self.hidden_size])
        if self.dropout is not None and self.dropout > 0.0:
            x_emb = fluid.layers.dropout(
                x_emb,
                dropout_prob=self.drop_out,
                dropout_implementation='upscale_in_train')
        rnn_out, last_hidden, last_cell = self.simple_lstm_rnn(x_emb, init_h,
                                                               init_c)
        rnn_out = fluid.layers.reshape(
            rnn_out, shape=[-1, self.num_steps, self.hidden_size])
201
        projection = fluid.layers.matmul(rnn_out, self.softmax_weight)
J
JiabinYang 已提交
202 203 204 205 206 207 208 209 210 211 212 213 214 215 216
        projection = fluid.layers.elementwise_add(projection, self.softmax_bias)
        projection = fluid.layers.reshape(
            projection, shape=[-1, self.vocab_size])
        projection = fluid.layers.reshape(
            projection, shape=[-1, self.vocab_size])
        loss = fluid.layers.softmax_with_cross_entropy(
            logits=projection, label=label, soft_label=False)
        loss = fluid.layers.reshape(loss, shape=[-1, self.num_steps])
        loss = fluid.layers.reduce_mean(loss, dim=[0])
        loss = fluid.layers.reduce_sum(loss)
        loss.permissions = True

        return loss, last_hidden, last_cell


L
lujun 已提交
217
class TestDygraphPtbRnn(unittest.TestCase):
218
    def test_ptb_rnn_cpu_float32(self):
J
JiabinYang 已提交
219 220 221 222 223 224 225 226
        seed = 90
        hidden_size = 10
        vocab_size = 1000
        num_layers = 1
        num_steps = 3
        init_scale = 0.1
        batch_size = 4

L
lujun 已提交
227
        with fluid.dygraph.guard():
J
JiabinYang 已提交
228 229 230 231
            fluid.default_startup_program().random_seed = seed
            fluid.default_main_program().random_seed = seed
            # TODO: marsyang1993 Change seed to
            ptb_model = PtbModel(
X
Xin Pan 已提交
232
                "ptb_model",
J
JiabinYang 已提交
233 234 235 236 237 238 239
                hidden_size=hidden_size,
                vocab_size=vocab_size,
                num_layers=num_layers,
                num_steps=num_steps,
                init_scale=init_scale)

            sgd = SGDOptimizer(learning_rate=1e-3)
240 241
            dy_param_updated = dict()
            dy_param_init = dict()
J
JiabinYang 已提交
242 243 244
            dy_loss = None
            last_hidden = None
            last_cell = None
245
            batch_num = 200
246 247

            for i in range(batch_num):
J
JiabinYang 已提交
248 249 250 251 252 253 254 255 256 257 258 259 260 261 262
                x_data = np.arange(12).reshape(4, 3).astype('int64')
                y_data = np.arange(1, 13).reshape(4, 3).astype('int64')
                x_data = x_data.reshape((-1, num_steps, 1))
                y_data = y_data.reshape((-1, 1))
                init_hidden_data = np.zeros(
                    (num_layers, batch_size, hidden_size), dtype='float32')
                init_cell_data = np.zeros(
                    (num_layers, batch_size, hidden_size), dtype='float32')
                x = to_variable(x_data)
                y = to_variable(y_data)
                init_hidden = to_variable(init_hidden_data)
                init_cell = to_variable(init_cell_data)
                dy_loss, last_hidden, last_cell = ptb_model(x, y, init_hidden,
                                                            init_cell)
                if i == 0:
263
                    for param in ptb_model.parameters():
J
JiabinYang 已提交
264 265 266
                        dy_param_init[param.name] = param._numpy()
                dy_loss._backward()
                sgd.minimize(dy_loss)
267 268 269 270
                ptb_model.clear_gradients()
                if i == batch_num - 1:
                    for param in ptb_model.parameters():
                        dy_param_updated[param.name] = param._numpy()
271 272 273 274 275

        with new_program_scope():
            fluid.default_startup_program().random_seed = seed
            fluid.default_main_program().random_seed = seed
            ptb_model = PtbModel(
X
Xin Pan 已提交
276
                "ptb_model",
277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297
                hidden_size=hidden_size,
                vocab_size=vocab_size,
                num_layers=num_layers,
                num_steps=num_steps,
                init_scale=init_scale)

            exe = fluid.Executor(fluid.CPUPlace())
            sgd = SGDOptimizer(learning_rate=1e-3)
            x = fluid.layers.data(name="x", shape=[-1, 3, 1], dtype='int64')
            y = fluid.layers.data(name="y", shape=[-1, 1], dtype='float32')
            init_hidden = fluid.layers.data(
                name="init_hidden", shape=[1], dtype='float32')
            init_cell = fluid.layers.data(
                name="init_cell", shape=[1], dtype='float32')

            static_loss, static_last_hidden, static_last_cell = ptb_model(
                x, y, init_hidden, init_cell)
            sgd.minimize(static_loss)
            static_param_updated = dict()
            static_param_init = dict()
            static_param_name_list = list()
298
            for param in ptb_model.parameters():
299 300 301 302 303 304
                static_param_name_list.append(param.name)

            out = exe.run(framework.default_startup_program(),
                          fetch_list=static_param_name_list)
            for i in range(len(static_param_name_list)):
                static_param_init[static_param_name_list[i]] = out[i]
J
JiabinYang 已提交
305 306 307
            static_loss_value = None
            static_last_cell_value = None
            static_last_hidden_value = None
308
            for i in range(batch_num):
309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327
                x_data = np.arange(12).reshape(4, 3).astype('int64')
                y_data = np.arange(1, 13).reshape(4, 3).astype('int64')
                x_data = x_data.reshape((-1, num_steps, 1))
                y_data = y_data.reshape((-1, 1))
                init_hidden_data = np.zeros(
                    (num_layers, batch_size, hidden_size), dtype='float32')
                init_cell_data = np.zeros(
                    (num_layers, batch_size, hidden_size), dtype='float32')
                fetch_list = [static_loss, static_last_hidden, static_last_cell]
                fetch_list.extend(static_param_name_list)
                out = exe.run(fluid.default_main_program(),
                              feed={
                                  "x": x_data,
                                  "y": y_data,
                                  "init_hidden": init_hidden_data,
                                  "init_cell": init_cell_data
                              },
                              fetch_list=fetch_list)
                static_loss_value = out[0]
328 329
                static_last_hidden_value = out[1]
                static_last_cell_value = out[2]
J
JiabinYang 已提交
330

331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346
                if i == batch_num - 1:
                    for k in range(3, len(out)):
                        static_param_updated[static_param_name_list[k -
                                                                    3]] = out[k]

        self.assertTrue(np.allclose(static_loss_value, dy_loss._numpy()))
        self.assertTrue(np.allclose(static_last_cell_value, last_cell._numpy()))
        self.assertTrue(
            np.allclose(static_last_hidden_value, last_hidden._numpy()))
        for key, value in six.iteritems(static_param_init):
            # print("static_init name: {}, value {}".format(key, value))
            # print("dy_init name: {}, value {}".format(key, dy_param_init[key]))
            self.assertTrue(np.allclose(value, dy_param_init[key], atol=1e-5))
        for key, value in six.iteritems(static_param_updated):
            # print("static name: {}, value {}".format(key, value))
            # print("dy name: {}, value {}".format(key, dy_param_updated[key]))
347
            self.assertTrue(
348 349
                np.allclose(
                    value, dy_param_updated[key], atol=1e-5))
J
JiabinYang 已提交
350 351 352 353


if __name__ == '__main__':
    unittest.main()