notest_rnn_encoder_decoer.py 9.9 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
#
# 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.

import numpy as np
import paddle.v2 as paddle
17 18 19 20
import paddle.fluid as fluid
import paddle.fluid.core as core
import paddle.fluid.framework as framework
import paddle.fluid.layers as layers
21
import contextlib
22 23
import math
import sys
24
import unittest
25
from paddle.fluid.executor import Executor
26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55

dict_size = 30000
source_dict_dim = target_dict_dim = dict_size
src_dict, trg_dict = paddle.dataset.wmt14.get_dict(dict_size)
hidden_dim = 32
embedding_dim = 16
batch_size = 10
max_length = 50
topk_size = 50
encoder_size = decoder_size = hidden_dim
IS_SPARSE = True
USE_PEEPHOLES = False


def bi_lstm_encoder(input_seq, hidden_size):
    input_forward_proj = fluid.layers.fc(input=input_seq,
                                         size=hidden_size * 4,
                                         bias_attr=True)
    forward, _ = fluid.layers.dynamic_lstm(
        input=input_forward_proj,
        size=hidden_size * 4,
        use_peepholes=USE_PEEPHOLES)
    input_backward_proj = fluid.layers.fc(input=input_seq,
                                          size=hidden_size * 4,
                                          bias_attr=True)
    backward, _ = fluid.layers.dynamic_lstm(
        input=input_backward_proj,
        size=hidden_size * 4,
        is_reverse=True,
        use_peepholes=USE_PEEPHOLES)
P
peterzhang2029 已提交
56 57 58 59 60

    forward_last = fluid.layers.sequence_last_step(input=forward)
    backward_first = fluid.layers.sequence_first_step(input=backward)

    return forward_last, backward_first
61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125


# FIXME(peterzhang2029): Replace this function with the lstm_unit_op.
def lstm_step(x_t, hidden_t_prev, cell_t_prev, size):
    def linear(inputs):
        return fluid.layers.fc(input=inputs, size=size, bias_attr=True)

    forget_gate = fluid.layers.sigmoid(x=linear([hidden_t_prev, x_t]))
    input_gate = fluid.layers.sigmoid(x=linear([hidden_t_prev, x_t]))
    output_gate = fluid.layers.sigmoid(x=linear([hidden_t_prev, x_t]))
    cell_tilde = fluid.layers.tanh(x=linear([hidden_t_prev, x_t]))

    cell_t = fluid.layers.sums(input=[
        fluid.layers.elementwise_mul(
            x=forget_gate, y=cell_t_prev), fluid.layers.elementwise_mul(
                x=input_gate, y=cell_tilde)
    ])

    hidden_t = fluid.layers.elementwise_mul(
        x=output_gate, y=fluid.layers.tanh(x=cell_t))

    return hidden_t, cell_t


def lstm_decoder_without_attention(target_embedding, decoder_boot, context,
                                   decoder_size):
    rnn = fluid.layers.DynamicRNN()

    cell_init = fluid.layers.fill_constant_batch_size_like(
        input=decoder_boot,
        value=0.0,
        shape=[-1, decoder_size],
        dtype='float32')
    cell_init.stop_gradient = False

    with rnn.block():
        current_word = rnn.step_input(target_embedding)
        context = rnn.static_input(context)

        hidden_mem = rnn.memory(init=decoder_boot, need_reorder=True)
        cell_mem = rnn.memory(init=cell_init)
        decoder_inputs = fluid.layers.concat(
            input=[context, current_word], axis=1)
        h, c = lstm_step(decoder_inputs, hidden_mem, cell_mem, decoder_size)
        rnn.update_memory(hidden_mem, h)
        rnn.update_memory(cell_mem, c)
        out = fluid.layers.fc(input=h,
                              size=target_dict_dim,
                              bias_attr=True,
                              act='softmax')
        rnn.output(out)
    return rnn()


def seq_to_seq_net():
    """Construct a seq2seq network."""

    src_word_idx = fluid.layers.data(
        name='source_sequence', shape=[1], dtype='int64', lod_level=1)

    src_embedding = fluid.layers.embedding(
        input=src_word_idx,
        size=[source_dict_dim, embedding_dim],
        dtype='float32')

P
peterzhang2029 已提交
126
    src_forward_last, src_backward_first = bi_lstm_encoder(
127 128
        input_seq=src_embedding, hidden_size=encoder_size)

P
peterzhang2029 已提交
129 130
    encoded_vector = fluid.layers.concat(
        input=[src_forward_last, src_backward_first], axis=1)
131

P
peterzhang2029 已提交
132
    decoder_boot = fluid.layers.fc(input=src_backward_first,
133 134 135 136 137 138 139 140 141 142 143 144 145
                                   size=decoder_size,
                                   bias_attr=False,
                                   act='tanh')

    trg_word_idx = fluid.layers.data(
        name='target_sequence', shape=[1], dtype='int64', lod_level=1)

    trg_embedding = fluid.layers.embedding(
        input=trg_word_idx,
        size=[target_dict_dim, embedding_dim],
        dtype='float32')

    prediction = lstm_decoder_without_attention(trg_embedding, decoder_boot,
P
peterzhang2029 已提交
146
                                                encoded_vector, decoder_size)
147 148 149 150 151
    label = fluid.layers.data(
        name='label_sequence', shape=[1], dtype='int64', lod_level=1)
    cost = fluid.layers.cross_entropy(input=prediction, label=label)
    avg_cost = fluid.layers.mean(x=cost)

K
Kexin Zhao 已提交
152
    return avg_cost, prediction
153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169


def to_lodtensor(data, place):
    seq_lens = [len(seq) for seq in data]
    cur_len = 0
    lod = [cur_len]
    for l in seq_lens:
        cur_len += l
        lod.append(cur_len)
    flattened_data = np.concatenate(data, axis=0).astype("int64")
    flattened_data = flattened_data.reshape([len(flattened_data), 1])
    res = core.LoDTensor()
    res.set(flattened_data, place)
    res.set_lod([lod])
    return res


170 171 172 173 174 175 176 177 178
def create_random_lodtensor(lod, place, low, high):
    data = np.random.random_integers(low, high, [lod[-1], 1]).astype("int64")
    res = fluid.LoDTensor()
    res.set(data, place)
    res.set_lod([lod])
    return res


def train(use_cuda, save_dirname=None):
K
Kexin Zhao 已提交
179
    [avg_cost, prediction] = seq_to_seq_net()
180 181 182 183 184 185 186 187 188

    optimizer = fluid.optimizer.Adagrad(learning_rate=1e-4)
    optimizer.minimize(avg_cost)

    train_data = paddle.batch(
        paddle.reader.shuffle(
            paddle.dataset.wmt14.train(dict_size), buf_size=1000),
        batch_size=batch_size)

189
    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
190 191 192 193 194 195 196 197 198 199
    exe = Executor(place)

    exe.run(framework.default_startup_program())

    batch_id = 0
    for pass_id in xrange(2):
        for data in train_data():
            word_data = to_lodtensor(map(lambda x: x[0], data), place)
            trg_word = to_lodtensor(map(lambda x: x[1], data), place)
            trg_word_next = to_lodtensor(map(lambda x: x[2], data), place)
200

201 202 203 204 205 206 207
            outs = exe.run(framework.default_main_program(),
                           feed={
                               'source_sequence': word_data,
                               'target_sequence': trg_word,
                               'label_sequence': trg_word_next
                           },
                           fetch_list=[avg_cost])
208

209 210 211
            avg_cost_val = np.array(outs[0])
            print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) +
                  " avg_cost=" + str(avg_cost_val))
212 213
            if math.isnan(float(avg_cost_val[0])):
                sys.exit("got NaN loss, training failed.")
214
            if batch_id > 3:
K
Kexin Zhao 已提交
215
                if save_dirname is not None:
216 217 218
                    fluid.io.save_inference_model(
                        save_dirname, ['source_sequence',
                                       'target_sequence'], [prediction], exe)
219 220
                return

221 222 223
            batch_id += 1


224
def infer(use_cuda, save_dirname=None):
K
Kexin Zhao 已提交
225 226 227
    if save_dirname is None:
        return

228
    place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
K
Kexin Zhao 已提交
229 230 231 232 233 234 235 236 237
    exe = fluid.Executor(place)

    # Use fluid.io.load_inference_model to obtain the inference program desc,
    # the feed_target_names (the names of variables that will be feeded 
    # data using feed operators), and the fetch_targets (variables that 
    # we want to obtain data from using fetch operators).
    [inference_program, feed_target_names,
     fetch_targets] = fluid.io.load_inference_model(save_dirname, exe)

238 239 240
    lod = [0, 4, 10]
    word_data = create_random_lodtensor(lod, place, low=0, high=1)
    trg_word = create_random_lodtensor(lod, place, low=0, high=1)
K
Kexin Zhao 已提交
241 242 243 244 245 246 247 248 249 250

    # Construct feed as a dictionary of {feed_target_name: feed_target_data}
    # and results will contain a list of data corresponding to fetch_targets.
    assert feed_target_names[0] == 'source_sequence'
    assert feed_target_names[1] == 'target_sequence'
    results = exe.run(inference_program,
                      feed={
                          feed_target_names[0]: word_data,
                          feed_target_names[1]: trg_word,
                      },
251 252 253 254 255 256
                      fetch_list=fetch_targets,
                      return_numpy=False)
    print(results[0].lod())
    np_data = np.array(results[0])
    print("Inference shape: ", np_data.shape)
    print("Inference results: ", np_data)
K
Kexin Zhao 已提交
257 258


259 260 261 262 263
def main(use_cuda):
    if use_cuda and not fluid.core.is_compiled_with_cuda():
        return

    # Directory for saving the trained model
K
Kexin Zhao 已提交
264
    save_dirname = "rnn_encoder_decoder.inference.model"
265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288

    train(use_cuda, save_dirname)
    infer(use_cuda, save_dirname)


class TestRnnEncoderDecoder(unittest.TestCase):
    def test_cuda(self):
        with self.scope_prog_guard():
            main(use_cuda=True)

    def test_cpu(self):
        with self.scope_prog_guard():
            main(use_cuda=False)

    @contextlib.contextmanager
    def scope_prog_guard(self):
        prog = fluid.Program()
        startup_prog = fluid.Program()
        scope = fluid.core.Scope()
        with fluid.scope_guard(scope):
            with fluid.program_guard(prog, startup_prog):
                yield


289
if __name__ == '__main__':
290
    unittest.main()