ELMO

PaddleNLP/ELMO/README.md

+<h1 align="center">ELMO</h1>
+
+## 介绍
+ELMO(Embeddings from Language Models)是一种新型深度语境化词表征，可对词进行复杂特征(如句法和语义)和词在语言语境中的变化进行建模(即对多义词进行建模)。PaddlePaddle版本该模型支持多卡训练，训练速度比主流实现快约1倍,  验证在中文词法分析任务上f1值提升0.68%。
+
+ELMO在大语料上以language model为训练目标，训练出bidirectional LSTM模型，利用LSTM产生词语的表征, 对下游NLP任务(如问答、分类、命名实体识别等）进行微调。
+
+## 基本配置及第三方安装包
+
+Python==2.7
+
+PaddlePaddle lastest版本
+
+numpy ==1.15.1
+
+six==1.11.0
+
+glob
+
+
+## 预训练模型
+
+1. 把文档文件切分成句子，并基于词表（参考vocabulary_min5k.txt）对句子进行切词。把文件切分成训练集trainset和测试集testset。
+
+```
+本 书 介绍 了 中国 经济 发展 的 内外 平衡 问题 、 亚洲 金融 危机 十 周年 回顾 与 反思 、 实践 中 的 城乡 统筹 发展 、 未来 十 年 中国 需要 研究 的 重大 课题 、 科学 发展 与 新型 工业 化 等 方面 。
+```
+```
+吴 敬 琏 曾经 提出 中国 股市 “ 赌场 论 ” ， 主张 维护 市场 规则 ， 保护 草根 阶层 生计 ， 被 誉 为 “ 中国 经济 学界 良心 ” ， 是 媒体 和 公众 眼中 的 学术 明星 
+```
+
+2. 训练模型
+
+```shell
+sh run.sh
+```
+
+3. 把checkpoint结果写入文件中。
+
+
+## 单机多卡训练
+
+模型支持单机多卡训练，需要在run.sh里export CUDA_VISIBLE_DEVICES设置指定卡,如下所示：
+```shell
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+```
+
+## 如何利用ELMO做微调
+
+1. 下载ELMO Paddle官方发布Checkout文件
+
+[PaddlePaddle官方发布Checkout文件下载地址](https://dureader.gz.bcebos.com/elmo/elmo_chinese_checkpoint.tar.gz)
+
+2. 在train部分中加载ELMO checkpoint文件
+
+```shell
+src_pretrain_model_path = '490001' #490001为ELMO checkpoint文件
+def if_exist(var):
+    path = os.path.join(src_pretrain_model_path, var.name)
+    exist = os.path.exists(path)
+    if exist:
+        print('Load model: %s' % path)
+    return exist
+
+fluid.io.load_vars(executor=exe, dirname=src_pretrain_model_path, predicate=if_exist, main_program=main_program) 
+```
+
+3. 在下游NLP任务文件夹中加入bilm.py文件
+
+4. 基于elmo词表（参考vocabulary_min5k.txt）对输入的句子或段落进行切词，并把切词的词转化为id,放入feed_dict中。
+
+5. 在下游任务网络定义中embedding部分加入ELMO网络的定义
+
+```shell
+#引入 bilm.py embedding部分和encoder部分
+from bilm import elmo_encoder
+from bilm import emb
+
+#word为输入elmo部分切词后的字典
+elmo_embedding = emb(word)
+elmo_enc= elmo_encoder(elmo_embedding)
+
+#与NLP任务中生成词向量word_embedding做连接操作
+word_embedding=layers.concat(input=[elmo_enc, word_embedding], axis=1)
+
+```
+
+
+## 参考论文
+[Deep contextualized word representations](https://arxiv.org/abs/1802.05365)

PaddleNLP/ELMO/args.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+# 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 absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import distutils.util
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument(
+        "--load_dir",
+        type=str,
+        default="",
+        help="Specify the path to load trained models.")
+    parser.add_argument(
+        "--load_pretraining_params",
+        type=str,
+        default="",
+        help="Specify the path to load pretrained model parameters, NOT including moment and learning_rate")
+    parser.add_argument(
+        "--batch_size",
+        type=int,
+        default=128,
+        help="The sequence number of a mini-batch data. (default: %(default)d)")
+    parser.add_argument(
+        "--embed_size",
+        type=int,
+        default=512,
+        help="The dimension of embedding table. (default: %(default)d)")
+    parser.add_argument(
+        "--hidden_size",
+        type=int,
+        default=4096,
+        help="The size of rnn hidden unit. (default: %(default)d)")
+    parser.add_argument(
+        "--num_layers",
+        type=int,
+        default=2,
+        help="The size of rnn layers. (default: %(default)d)")
+    parser.add_argument(
+        "--num_steps",
+        type=int,
+        default=20,
+        help="The size of sequence len. (default: %(default)d)")
+    parser.add_argument(
+        "--data_path", type=str, help="all the data for train,valid,test")
+    parser.add_argument("--vocab_path", type=str, help="vocab file path")
+    parser.add_argument(
+        '--use_gpu', type=bool, default=False, help='whether using gpu')
+    parser.add_argument('--enable_ce', action='store_true')
+    parser.add_argument('--test_nccl', action='store_true')
+    parser.add_argument('--optim', default='adagrad', help='optimizer type')
+    parser.add_argument('--sample_softmax', action='store_true')
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=0.2,
+        help="Learning rate used to train the model. (default: %(default)f)")
+    parser.add_argument(
+        "--log_interval",
+        type=int,
+        default=100,
+        help="log the train loss every n batches."
+        "(default: %(default)d)")
+    parser.add_argument(
+        "--save_interval",
+        type=int,
+        default=10000,
+        help="log the train loss every n batches."
+        "(default: %(default)d)")
+    parser.add_argument(
+        "--dev_interval",
+        type=int,
+        default=10000,
+        help="cal dev loss every n batches."
+        "(default: %(default)d)")
+    parser.add_argument('--dropout', type=float, default=0.1)
+    parser.add_argument('--max_grad_norm', type=float, default=10.0)
+    parser.add_argument('--proj_clip', type=float, default=3.0)
+    parser.add_argument('--cell_clip', type=float, default=3.0)
+    parser.add_argument('--max_epoch', type=float, default=10)
+    parser.add_argument('--local', type=bool, default=False)
+    parser.add_argument('--shuffle', type=bool, default=False)
+    parser.add_argument('--use_custom_samples', type=bool, default=False)
+    parser.add_argument('--para_save_dir', type=str, default='model_new')
+    parser.add_argument('--train_path', type=str, default='')
+    parser.add_argument('--test_path', type=str, default='')
+    parser.add_argument('--update_method', type=str, default='nccl2')
+    parser.add_argument('--random_seed', type=int, default=0)
+    parser.add_argument('--n_negative_samples_batch', type=int, default=8000)
+    args = parser.parse_args()
+
+    return args

PaddleNLP/ELMO/bilm.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+# 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.
+
+# This file is used to finetone
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import numpy
+import paddle.fluid.layers as layers
+import paddle.fluid as fluid
+import numpy as np
+
+# if you use our release weight layers,do not use the args.
+cell_clip = 3.0
+proj_clip = 3.0
+hidden_size = 4096
+vocab_size = 52445
+embed_size = 512
+# according to orginal paper, dropout need to be modifyed on finetone
+modify_dropout = 1
+proj_size = 512
+num_layers = 2
+random_seed = 0
+dropout_rate = 0.5
+
+
+def dropout(input):
+    return layers.dropout(
+        input,
+        dropout_prob=dropout_rate,
+        dropout_implementation="upscale_in_train",
+        seed=random_seed,
+        is_test=False)
+
+def lstmp_encoder(input_seq, gate_size, h_0, c_0, para_name):
+    # A lstm encoder implementation with projection.
+    # Linear transformation part for input gate, output gate, forget gate
+    # and cell activation vectors need be done outside of dynamic_lstm.
+    # So the output size is 4 times of gate_size.
+
+    input_proj = layers.fc(input=input_seq,
+                           param_attr=fluid.ParamAttr(
+                               name=para_name + '_gate_w', initializer=init),
+                           size=gate_size * 4,
+                           act=None,
+                           bias_attr=False)
+    hidden, cell = layers.dynamic_lstmp(
+        input=input_proj,
+        size=gate_size * 4,
+        proj_size=proj_size,
+        h_0=h_0,
+        c_0=c_0,
+        use_peepholes=False,
+        proj_clip=proj_clip,
+        cell_clip=cell_clip,
+        proj_activation="identity",
+        param_attr=fluid.ParamAttr(initializer=None),
+        bias_attr=fluid.ParamAttr(initializer=None))
+    return hidden, cell, input_proj
+
+
+def encoder(x_emb,
+            init_hidden=None,
+            init_cell=None,
+            para_name=''):
+    rnn_input = x_emb
+    rnn_outs = []
+    rnn_outs_ori = []
+    cells = []
+    projs = []
+    for i in range(num_layers):
+        if init_hidden and init_cell:
+            h0 = layers.squeeze(
+                layers.slice(
+                    init_hidden, axes=[0], starts=[i], ends=[i + 1]),
+                axes=[0])
+            c0 = layers.squeeze(
+                layers.slice(
+                    init_cell, axes=[0], starts=[i], ends=[i + 1]),
+                axes=[0])
+        else:
+            h0 = c0 = None
+        rnn_out, cell, input_proj = lstmp_encoder(
+            rnn_input, hidden_size, h0, c0,
+            para_name + 'layer{}'.format(i + 1))
+        rnn_out_ori = rnn_out
+        if i > 0:
+            rnn_out = rnn_out + rnn_input
+        rnn_out.stop_gradient = True
+        rnn_outs.append(rnn_out)
+        rnn_outs_ori.append(rnn_out_ori)
+    # add weight layers for finetone
+    a1 = layers.create_parameter(
+        [1], dtype="float32", name="gamma1")
+    a2 = layers.create_parameter(
+        [1], dtype="float32", name="gamma2")
+    rnn_outs[0].stop_gradient = True
+    rnn_outs[1].stop_gradient = True
+    num_layer1 = rnn_outs[0] * a1
+    num_layer2 = rnn_outs[1] * a2
+    output_layer = num_layer1 * 0.5 + num_layer2 * 0.5
+    return output_layer, rnn_outs_ori
+
+
+def emb(x):
+    x_emb = layers.embedding(
+        input=x,
+        size=[vocab_size, embed_size],
+        dtype='float32',
+        is_sparse=False,
+        param_attr=fluid.ParamAttr(name='embedding_para'))
+    return x_emb
+
+
+def elmo_encoder(x_emb):
+    x_emb_r = fluid.layers.sequence_reverse(x_emb, name=None)
+    fw_hiddens, fw_hiddens_ori = encoder(
+        x_emb,
+        para_name='fw_')
+    bw_hiddens, bw_hiddens_ori = encoder(
+        x_emb_r,
+        para_name='bw_')
+    embedding = layers.concat(input=[fw_hiddens, bw_hiddens], axis=1)
+    # add dropout on finetone
+    embedding = dropout(embedding)
+    a = layers.create_parameter(
+        [1], dtype="float32", name="gamma")
+    embedding = embedding * a
+    return embedding

PaddleNLP/ELMO/data.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+# 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 glob
+import random
+
+import numpy as np
+import io
+import six
+
+class Vocabulary(object):
+    '''
+    A token vocabulary.  Holds a map from token to ids and provides
+    a method for encoding text to a sequence of ids.
+    '''
+
+    def __init__(self, filename, validate_file=False):
+        '''
+        filename = the vocabulary file.  It is a flat text file with one
+            (normalized) token per line.  In addition, the file should also
+            contain the special tokens <S>, </S>, <UNK> (case sensitive).
+        '''
+        self._id_to_word = []
+        self._word_to_id = {}
+        self._unk = -1
+        self._bos = -1
+        self._eos = -1
+
+        with io.open(filename, 'r', encoding='utf-8') as f:
+            idx = 0
+            for line in f:
+                word_name = line.strip()
+                if word_name == '<S>':
+                    self._bos = idx
+                elif word_name == '</S>':
+                    self._eos = idx
+                elif word_name == '<UNK>':
+                    self._unk = idx
+                if word_name == '!!!MAXTERMID':
+                    continue
+
+                self._id_to_word.append(word_name)
+                self._word_to_id[word_name] = idx
+                idx += 1
+
+        # check to ensure file has special tokens
+        if validate_file:
+            if self._bos == -1 or self._eos == -1 or self._unk == -1:
+                raise ValueError("Ensure the vocabulary file has "
+                                 "<S>, </S>, <UNK> tokens")
+
+    @property
+    def bos(self):
+        return self._bos
+
+    @property
+    def eos(self):
+        return self._eos
+
+    @property
+    def unk(self):
+        return self._unk
+
+    @property
+    def size(self):
+        return len(self._id_to_word)
+
+    def word_to_id(self, word):
+        if word in self._word_to_id:
+            return self._word_to_id[word]
+        return self.unk
+
+    def id_to_word(self, cur_id):
+        return self._id_to_word[cur_id]
+
+    def decode(self, cur_ids):
+        """Convert a list of ids to a sentence, with space inserted."""
+        return ' '.join([self.id_to_word(cur_id) for cur_id in cur_ids])
+
+    def encode(self, sentence, reverse=False, split=True):
+        """Convert a sentence to a list of ids, with special tokens added.
+        Sentence is a single string with tokens separated by whitespace.
+
+        If reverse, then the sentence is assumed to be reversed, and
+            this method will swap the BOS/EOS tokens appropriately."""
+
+        if split:
+            word_ids = [
+                self.word_to_id(cur_word) for cur_word in sentence.split()
+            ]
+        else:
+            word_ids = [self.word_to_id(cur_word) for cur_word in sentence]
+
+        if reverse:
+            return np.array([self.eos] + word_ids + [self.bos], dtype=np.int32)
+        else:
+            return np.array([self.bos] + word_ids + [self.eos], dtype=np.int32)
+
+
+class UnicodeCharsVocabulary(Vocabulary):
+    """Vocabulary containing character-level and word level information.
+
+    Has a word vocabulary that is used to lookup word ids and
+    a character id that is used to map words to arrays of character ids.
+
+    The character ids are defined by ord(c) for c in word.encode('utf-8')
+    This limits the total number of possible char ids to 256.
+    To this we add 5 additional special ids: begin sentence, end sentence,
+        begin word, end word and padding.
+
+    WARNING: for prediction, we add +1 to the output ids from this
+    class to create a special padding id (=0).  As a result, we suggest
+    you use the `Batcher`, `TokenBatcher`, and `LMDataset` classes instead
+    of this lower level class.  If you are using this lower level class,
+    then be sure to add the +1 appropriately, otherwise embeddings computed
+    from the pre-trained model will be useless.
+    """
+
+    def __init__(self, filename, max_word_length, **kwargs):
+        super(UnicodeCharsVocabulary, self).__init__(filename, **kwargs)
+        self._max_word_length = max_word_length
+
+        # char ids 0-255 come from utf-8 encoding bytes
+        # assign 256-300 to special chars
+        self.bos_char = 256  # <begin sentence>
+        self.eos_char = 257  # <end sentence>
+        self.bow_char = 258  # <begin word>
+        self.eow_char = 259  # <end word>
+        self.pad_char = 260  # <padding>
+
+        num_words = len(self._id_to_word)
+
+        self._word_char_ids = np.zeros(
+            [num_words, max_word_length], dtype=np.int32)
+
+        # the charcter representation of the begin/end of sentence characters
+        def _make_bos_eos(c):
+            r = np.zeros([self.max_word_length], dtype=np.int32)
+            r[:] = self.pad_char
+            r[0] = self.bow_char
+            r[1] = c
+            r[2] = self.eow_char
+            return r
+
+        self.bos_chars = _make_bos_eos(self.bos_char)
+        self.eos_chars = _make_bos_eos(self.eos_char)
+
+        for i, word in enumerate(self._id_to_word):
+            self._word_char_ids[i] = self._convert_word_to_char_ids(word)
+
+        self._word_char_ids[self.bos] = self.bos_chars
+        self._word_char_ids[self.eos] = self.eos_chars
+
+    @property
+    def word_char_ids(self):
+        return self._word_char_ids
+
+    @property
+    def max_word_length(self):
+        return self._max_word_length
+
+    def _convert_word_to_char_ids(self, word):
+        code = np.zeros([self.max_word_length], dtype=np.int32)
+        code[:] = self.pad_char
+
+        word_encoded = word.encode('utf-8',
+                                   'ignore')[:(self.max_word_length - 2)]
+        code[0] = self.bow_char
+        for k, chr_id in enumerate(word_encoded, start=1):
+            code[k] = ord(chr_id)
+        code[k + 1] = self.eow_char
+
+        return code
+
+    def word_to_char_ids(self, word):
+        if word in self._word_to_id:
+            return self._word_char_ids[self._word_to_id[word]]
+        else:
+            return self._convert_word_to_char_ids(word)
+
+    def encode_chars(self, sentence, reverse=False, split=True):
+        '''
+        Encode the sentence as a white space delimited string of tokens.
+        '''
+        if split:
+            chars_ids = [
+                self.word_to_char_ids(cur_word)
+                for cur_word in sentence.split()
+            ]
+        else:
+            chars_ids = [
+                self.word_to_char_ids(cur_word) for cur_word in sentence
+            ]
+        if reverse:
+            return np.vstack([self.eos_chars] + chars_ids + [self.bos_chars])
+        else:
+            return np.vstack([self.bos_chars] + chars_ids + [self.eos_chars])
+
+
+class Batcher(object):
+    ''' 
+    Batch sentences of tokenized text into character id matrices.
+    '''
+
+    # def __init__(self, lm_vocab_file: str, max_token_length: int):
+    def __init__(self, lm_vocab_file, max_token_length):
+        '''
+        lm_vocab_file = the language model vocabulary file (one line per
+            token)
+        max_token_length = the maximum number of characters in each token
+        '''
+        max_token_length = int(max_token_length)
+        self._lm_vocab = UnicodeCharsVocabulary(lm_vocab_file,
+                                                max_token_length)
+        self._max_token_length = max_token_length
+
+    # def batch_sentences(self, sentences: List[List[str]]):
+    def batch_sentences(self, sentences):
+        '''
+        Batch the sentences as character ids
+        Each sentence is a list of tokens without <s> or </s>, e.g.
+        [['The', 'first', 'sentence', '.'], ['Second', '.']]
+        '''
+        n_sentences = len(sentences)
+        max_length = max(len(sentence) for sentence in sentences) + 2
+
+        X_char_ids = np.zeros(
+            (n_sentences, max_length, self._max_token_length), dtype=np.int64)
+
+        for k, sent in enumerate(sentences):
+            length = len(sent) + 2
+            char_ids_without_mask = self._lm_vocab.encode_chars(
+                sent, split=False)
+            # add one so that 0 is the mask value
+            X_char_ids[k, :length, :] = char_ids_without_mask + 1
+
+        return X_char_ids
+
+
+class TokenBatcher(object):
+    ''' 
+    Batch sentences of tokenized text into token id matrices.
+    '''
+
+    def __init__(self, lm_vocab_file):
+        # def __init__(self, lm_vocab_file: str):
+        '''
+        lm_vocab_file = the language model vocabulary file (one line per
+            token)
+        '''
+        self._lm_vocab = Vocabulary(lm_vocab_file)
+
+    # def batch_sentences(self, sentences: List[List[str]]):
+    def batch_sentences(self, sentences):
+        '''
+        Batch the sentences as character ids
+        Each sentence is a list of tokens without <s> or </s>, e.g.
+        [['The', 'first', 'sentence', '.'], ['Second', '.']]
+        '''
+        n_sentences = len(sentences)
+        max_length = max(len(sentence) for sentence in sentences) + 2
+
+        X_ids = np.zeros((n_sentences, max_length), dtype=np.int64)
+
+        for k, sent in enumerate(sentences):
+            length = len(sent) + 2
+            ids_without_mask = self._lm_vocab.encode(sent, split=False)
+            # add one so that 0 is the mask value
+            X_ids[k, :length] = ids_without_mask + 1
+
+        return X_ids
+
+
+##### for training
+def _get_batch(generator, batch_size, num_steps, max_word_length):
+    """Read batches of input."""
+    cur_stream = [None] * batch_size
+
+    no_more_data = False
+    while True:
+        inputs = np.zeros([batch_size, num_steps], np.int32)
+        if max_word_length is not None:
+            char_inputs = np.zeros([batch_size, num_steps, max_word_length],
+                                   np.int32)
+        else:
+            char_inputs = None
+        targets = np.zeros([batch_size, num_steps], np.int32)
+
+        for i in range(batch_size):
+            cur_pos = 0
+
+            while cur_pos < num_steps:
+                if cur_stream[i] is None or len(cur_stream[i][0]) <= 1:
+                    try:
+                        cur_stream[i] = list(next(generator))
+                    except StopIteration:
+                        # No more data, exhaust current streams and quit
+                        no_more_data = True
+                        break
+
+                how_many = min(len(cur_stream[i][0]) - 1, num_steps - cur_pos)
+                next_pos = cur_pos + how_many
+
+                inputs[i, cur_pos:next_pos] = cur_stream[i][0][:how_many]
+                if max_word_length is not None:
+                    char_inputs[i, cur_pos:next_pos] = cur_stream[i][
+                        1][:how_many]
+                targets[i, cur_pos:next_pos] = cur_stream[i][0][1:how_many + 1]
+
+                cur_pos = next_pos
+
+                cur_stream[i][0] = cur_stream[i][0][how_many:]
+                if max_word_length is not None:
+                    cur_stream[i][1] = cur_stream[i][1][how_many:]
+
+        if no_more_data:
+            # There is no more data.  Note: this will not return data
+            # for the incomplete batch
+            break
+
+        X = {
+            'token_ids': inputs,
+            'tokens_characters': char_inputs,
+            'next_token_id': targets
+        }
+
+        yield X
+
+
+class LMDataset(object):
+    """
+    Hold a language model dataset.
+
+    A dataset is a list of tokenized files.  Each file contains one sentence
+        per line.  Each sentence is pre-tokenized and white space joined.
+    """
+
+    def __init__(self,
+                 filepattern,
+                 vocab,
+                 reverse=False,
+                 test=False,
+                 shuffle_on_load=False):
+        '''
+        filepattern = a glob string that specifies the list of files.
+        vocab = an instance of Vocabulary or UnicodeCharsVocabulary
+        reverse = if True, then iterate over tokens in each sentence in reverse
+        test = if True, then iterate through all data once then stop.
+            Otherwise, iterate forever.
+        shuffle_on_load = if True, then shuffle the sentences after loading.
+        '''
+        self._vocab = vocab
+        self._all_shards = glob.glob(filepattern)
+        print('Found %d shards at %s' % (len(self._all_shards), filepattern))
+        if test:
+            self._all_shards = list(np.random.choice(self._all_shards, size=4))
+            print('sampled %d shards at %s' % (len(self._all_shards), filepattern))
+        self._shards_to_choose = []
+
+        self._reverse = reverse
+        self._test = test
+        self._shuffle_on_load = shuffle_on_load
+        self._use_char_inputs = hasattr(vocab, 'encode_chars')
+
+        self._ids = self._load_random_shard()
+
+    def _choose_random_shard(self):
+        if len(self._shards_to_choose) == 0:
+            self._shards_to_choose = list(self._all_shards)
+            random.shuffle(self._shards_to_choose)
+        shard_name = self._shards_to_choose.pop()
+        return shard_name
+
+    def _load_random_shard(self):
+        """Randomly select a file and read it."""
+        if self._test:
+            if len(self._all_shards) == 0:
+                # we've loaded all the data
+                # this will propogate up to the generator in get_batch
+                # and stop iterating
+                raise StopIteration
+            else:
+                shard_name = self._all_shards.pop()
+        else:
+            # just pick a random shard
+            shard_name = self._choose_random_shard()
+
+        ids = self._load_shard(shard_name)
+        self._i = 0
+        self._nids = len(ids)
+        return ids
+
+    def _load_shard(self, shard_name):
+        """Read one file and convert to ids.
+
+        Args:
+            shard_name: file path.
+
+        Returns:
+            list of (id, char_id) tuples.
+        """
+        print('Loading data from: %s' % shard_name)
+        with io.open(shard_name, 'r', encoding='utf-8') as f:
+            sentences_raw = f.readlines()
+
+        if self._reverse:
+            sentences = []
+            for sentence in sentences_raw:
+                splitted = sentence.split()
+                splitted.reverse()
+                sentences.append(' '.join(splitted))
+        else:
+            sentences = sentences_raw
+
+        if self._shuffle_on_load:
+            print('shuffle sentences')
+            random.shuffle(sentences)
+
+        ids = [
+            self.vocab.encode(sentence, self._reverse)
+            for sentence in sentences
+        ]
+        if self._use_char_inputs:
+            chars_ids = [
+                self.vocab.encode_chars(sentence, self._reverse)
+                for sentence in sentences
+            ]
+        else:
+            chars_ids = [None] * len(ids)
+
+        print('Loaded %d sentences.' % len(ids))
+        print('Finished loading')
+        return list(zip(ids, chars_ids))
+
+    def get_sentence(self):
+        while True:
+            if self._i == self._nids:
+                self._ids = self._load_random_shard()
+            ret = self._ids[self._i]
+            self._i += 1
+            yield ret
+
+    @property
+    def max_word_length(self):
+        if self._use_char_inputs:
+            return self._vocab.max_word_length
+        else:
+            return None
+
+    def iter_batches(self, batch_size, num_steps):
+        for X in _get_batch(self.get_sentence(), batch_size, num_steps,
+                            self.max_word_length):
+
+            # token_ids = (batch_size, num_steps)
+            # char_inputs = (batch_size, num_steps, 50) of character ids
+            # targets = word ID of next word (batch_size, num_steps)
+            yield X
+
+    @property
+    def vocab(self):
+        return self._vocab
+
+
+class BidirectionalLMDataset(object):
+    def __init__(self, filepattern, vocab, test=False, shuffle_on_load=False):
+        '''
+        bidirectional version of LMDataset
+        '''
+        self._data_forward = LMDataset(
+            filepattern,
+            vocab,
+            reverse=False,
+            test=test,
+            shuffle_on_load=shuffle_on_load)
+        self._data_reverse = LMDataset(
+            filepattern,
+            vocab,
+            reverse=True,
+            test=test,
+            shuffle_on_load=shuffle_on_load)
+
+    def iter_batches(self, batch_size, num_steps):
+        max_word_length = self._data_forward.max_word_length
+
+        for X, Xr in six.moves.zip(
+                _get_batch(self._data_forward.get_sentence(), batch_size,
+                           num_steps, max_word_length),
+                _get_batch(self._data_reverse.get_sentence(), batch_size,
+                           num_steps, max_word_length)):
+
+            for k, v in Xr.items():
+                X[k + '_reverse'] = v
+
+            yield X
+
+
+class InvalidNumberOfCharacters(Exception):
+    pass

PaddleNLP/ELMO/lm_model.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+# 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 absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import paddle.fluid.layers as layers
+import paddle.fluid as fluid
+import numpy as np
+
+
+def dropout(input, test_mode, args):
+    if args.dropout and (not test_mode):
+        return layers.dropout(
+            input,
+            dropout_prob=args.dropout,
+            dropout_implementation="upscale_in_train",
+            seed=args.random_seed,
+            is_test=False)
+    else:
+        return input
+
+
+def lstmp_encoder(input_seq, gate_size, h_0, c_0, para_name, proj_size, test_mode, args):
+    # A lstm encoder implementation with projection.
+    # Linear transformation part for input gate, output gate, forget gate
+    # and cell activation vectors need be done outside of dynamic_lstm.
+    # So the output size is 4 times of gate_size.
+
+    input_seq = dropout(input_seq, test_mode, args)
+    input_proj = layers.fc(input=input_seq,
+                           param_attr=fluid.ParamAttr(
+                               name=para_name + '_gate_w', initializer=None),
+                           size=gate_size * 4,
+                           act=None,
+                           bias_attr=False)
+    hidden, cell = layers.dynamic_lstmp(
+        input=input_proj,
+        size=gate_size * 4,
+        proj_size=proj_size,
+        h_0=h_0,
+        c_0=c_0,
+        use_peepholes=False,
+        proj_clip=args.proj_clip,
+        cell_clip=args.cell_clip,
+        proj_activation="identity",
+        param_attr=fluid.ParamAttr(initializer=None),
+        bias_attr=fluid.ParamAttr(initializer=None))
+
+    return hidden, cell, input_proj
+
+
+def encoder(x,
+            y,
+            vocab_size,
+            emb_size,
+            init_hidden=None,
+            init_cell=None,
+            para_name='',
+            custom_samples=None,
+            custom_probabilities=None,
+            test_mode=False,
+            args=None):
+    x_emb = layers.embedding(
+        input=x,
+        size=[vocab_size, emb_size],
+        dtype='float32',
+        is_sparse=False,
+        param_attr=fluid.ParamAttr(name='embedding_para'))
+    rnn_input = x_emb
+    rnn_outs = []
+    rnn_outs_ori = []
+    cells = []
+    projs = []
+    for i in range(args.num_layers):
+        rnn_input = dropout(rnn_input, test_mode, args)
+        if init_hidden and init_cell:
+            h0 = layers.squeeze(
+                layers.slice(
+                    init_hidden, axes=[0], starts=[i], ends=[i + 1]),
+                axes=[0])
+            c0 = layers.squeeze(
+                layers.slice(
+                    init_cell, axes=[0], starts=[i], ends=[i + 1]),
+                axes=[0])
+        else:
+            h0 = c0 = None
+        rnn_out, cell, input_proj = lstmp_encoder(
+            rnn_input, args.hidden_size, h0, c0,
+            para_name + 'layer{}'.format(i + 1), emb_size, test_mode, args)
+        rnn_out_ori = rnn_out
+        if i > 0:
+            rnn_out = rnn_out + rnn_input
+        rnn_out = dropout(rnn_out, test_mode, args)
+        cell = dropout(cell, test_mode, args)
+        rnn_outs.append(rnn_out)
+        rnn_outs_ori.append(rnn_out_ori)
+        rnn_input = rnn_out
+        cells.append(cell)
+        projs.append(input_proj)
+
+    softmax_weight = layers.create_parameter(
+        [vocab_size, emb_size], dtype="float32", name="softmax_weight")
+    softmax_bias = layers.create_parameter(
+        [vocab_size], dtype="float32", name='softmax_bias')
+    projection = layers.matmul(rnn_outs[-1], softmax_weight, transpose_y=True)
+    projection = layers.elementwise_add(projection, softmax_bias)
+
+    projection = layers.reshape(projection, shape=[-1, vocab_size])
+
+    if args.sample_softmax and (not test_mode):
+        loss = layers.sampled_softmax_with_cross_entropy(
+            logits=projection,
+            label=y,
+            num_samples=args.n_negative_samples_batch,
+            seed=args.random_seed)
+    else:
+        label = layers.one_hot(input=y, depth=vocab_size)
+        loss = layers.softmax_with_cross_entropy(
+            logits=projection, label=label, soft_label=True)
+    return [x_emb, projection, loss], rnn_outs, rnn_outs_ori, cells, projs
+
+
+class LanguageModel(object):
+    def __init__(self, args, vocab_size, test_mode):
+        self.args = args
+        self.vocab_size = vocab_size
+        self.test_mode = test_mode
+
+    def build(self):
+        args = self.args
+        emb_size = args.embed_size
+        proj_size = args.embed_size
+        hidden_size = args.hidden_size
+        batch_size = args.batch_size
+        num_layers = args.num_layers
+        num_steps = args.num_steps
+
+        lstm_outputs = []
+
+        x_f = layers.data(name="x", shape=[1], dtype='int64', lod_level=1)
+        y_f = layers.data(name="y", shape=[1], dtype='int64', lod_level=1)
+
+        x_b = layers.data(name="x_r", shape=[1], dtype='int64', lod_level=1)
+        y_b = layers.data(name="y_r", shape=[1], dtype='int64', lod_level=1)
+
+        init_hiddens_ = layers.data(
+            name="init_hiddens", shape=[1], dtype='float32')
+        init_cells_ = layers.data(
+            name="init_cells", shape=[1], dtype='float32')
+
+        init_hiddens = layers.reshape(
+            init_hiddens_, shape=[2 * num_layers, -1, proj_size])
+        init_cells = layers.reshape(
+            init_cells_, shape=[2 * num_layers, -1, hidden_size])
+
+        init_hidden = layers.slice(
+            init_hiddens, axes=[0], starts=[0], ends=[num_layers])
+        init_cell = layers.slice(
+            init_cells, axes=[0], starts=[0], ends=[num_layers])
+        init_hidden_r = layers.slice(
+            init_hiddens, axes=[0], starts=[num_layers],
+            ends=[2 * num_layers])
+        init_cell_r = layers.slice(
+            init_cells, axes=[0], starts=[num_layers], ends=[2 * num_layers])
+
+        if args.use_custom_samples:
+            custom_samples = layers.data(
+                name="custom_samples",
+                shape=[args.n_negative_samples_batch + 1],
+                dtype='int64',
+                lod_level=1)
+            custom_samples_r = layers.data(
+                name="custom_samples_r",
+                shape=[args.n_negative_samples_batch + 1],
+                dtype='int64',
+                lod_level=1)
+            custom_probabilities = layers.data(
+                name="custom_probabilities",
+                shape=[args.n_negative_samples_batch + 1],
+                dtype='float32',
+                lod_level=1)
+        else:
+            custom_samples = None
+            custom_samples_r = None
+            custom_probabilities = None
+
+        forward, fw_hiddens, fw_hiddens_ori, fw_cells, fw_projs = encoder(
+            x_f,
+            y_f,
+            self.vocab_size,
+            emb_size,
+            init_hidden,
+            init_cell,
+            para_name='fw_',
+            custom_samples=custom_samples,
+            custom_probabilities=custom_probabilities,
+            test_mode=self.test_mode,
+            args=args)
+        backward, bw_hiddens, bw_hiddens_ori, bw_cells, bw_projs = encoder(
+            x_b,
+            y_b,
+            self.vocab_size,
+            emb_size,
+            init_hidden_r,
+            init_cell_r,
+            para_name='bw_',
+            custom_samples=custom_samples_r,
+            custom_probabilities=custom_probabilities,
+            test_mode=self.test_mode,
+            args=args)
+
+        losses = layers.concat([forward[-1], backward[-1]])
+        self.loss = layers.reduce_mean(losses)
+        self.loss.permissions = True
+        self.loss.persistable = True
+        self.grad_vars = [x_f, y_f, x_b, y_b, self.loss]
+        self.grad_vars_name = ['x', 'y', 'x_r', 'y_r', 'final_loss']
+        fw_vars_name = ['x_emb', 'proj', 'loss'] + [
+            'init_hidden', 'init_cell'
+        ] + ['rnn_out', 'rnn_out2', 'cell', 'cell2', 'xproj', 'xproj2']
+        bw_vars_name = ['x_emb_r', 'proj_r', 'loss_r'] + [
+            'init_hidden_r', 'init_cell_r'
+        ] + [
+            'rnn_out_r', 'rnn_out2_r', 'cell_r', 'cell2_r', 'xproj_r',
+            'xproj2_r'
+        ]
+        fw_vars = forward + [init_hidden, init_cell
+                             ] + fw_hiddens + fw_cells + fw_projs
+        bw_vars = backward + [init_hidden_r, init_cell_r
+                              ] + bw_hiddens + bw_cells + bw_projs
+        for i in range(len(fw_vars_name)):
+            self.grad_vars.append(fw_vars[i])
+            self.grad_vars.append(bw_vars[i])
+            self.grad_vars_name.append(fw_vars_name[i])
+            self.grad_vars_name.append(bw_vars_name[i])
+        if args.use_custom_samples:
+            self.feed_order = [
+                'x', 'y', 'x_r', 'y_r', 'custom_samples', 'custom_samples_r',
+                'custom_probabilities'
+            ]
+        else:
+            self.feed_order = ['x', 'y', 'x_r', 'y_r']
+        self.last_hidden = [
+            fluid.layers.sequence_last_step(input=x)
+            for x in fw_hiddens_ori + bw_hiddens_ori
+        ]
+        self.last_cell = [
+            fluid.layers.sequence_last_step(input=x)
+            for x in fw_cells + bw_cells
+        ]
+        self.last_hidden = layers.concat(self.last_hidden, axis=0)
+        self.last_hidden.persistable = True
+        self.last_cell = layers.concat(self.last_cell, axis=0)
+        self.last_cell.persistable = True

PaddleNLP/ELMO/reader.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+# 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 absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import os
+import sys
+import numpy as np
+
+Py3 = sys.version_info[0] == 3
+
+
+def listDir(rootDir):
+    res = []
+    for filename in os.listdir(rootDir):
+        pathname = os.path.join(rootDir, filename)
+        if (os.path.isfile(pathname)):
+            res.append(pathname)
+    return res
+
+
+_unk = -1
+_bos = -1
+_eos = -1
+
+
+def _read_words(filename):
+    data = []
+    with open(filename, "r") as f:
+        return f.read().decode("utf-8").replace("\n", "<eos>").split()
+
+
+def _build_vocab(filename):
+    data = _read_words(filename)
+
+    counter = collections.Counter(data)
+    count_pairs = sorted(counter.items(), key=lambda x: (-x[1], x[0]))
+
+    words, _ = list(zip(*count_pairs))
+
+    print("vocab word num", len(words))
+    word_to_id = dict(zip(words, range(len(words))))
+
+    return word_to_id
+
+
+def _load_vocab(filename):
+    with open(filename, "r") as f:
+        words = f.read().decode("utf-8").replace("\n", " ").split()
+        word_to_id = dict(zip(words, range(len(words))))
+        _unk = word_to_id['<S>']
+        _eos = word_to_id['</S>']
+        _unk = word_to_id['<UNK>']
+        return word_to_id
+
+
+def _file_to_word_ids(filenames, word_to_id):
+    for filename in filenames:
+        data = _read_words(filename)
+        for id in [word_to_id[word] for word in data if word in word_to_id]:
+            yield id
+
+
+def ptb_raw_data(data_path=None, vocab_path=None, args=None):
+    """Load PTB raw data from data directory "data_path".
+
+  Reads PTB text files, converts strings to integer ids,
+  and performs mini-batching of the inputs.
+
+  The PTB dataset comes from Tomas Mikolov's webpage:
+
+  http://www.fit.vutbr.cz/~imikolov/rnnlm/simple-examples.tgz
+
+  Args:
+    data_path: string path to the directory where simple-examples.tgz has
+      been extracted.
+
+  Returns:
+    tuple (train_data, valid_data, test_data, vocabulary)
+    where each of the data objects can be passed to PTBIterator.
+  """
+    if vocab_path:
+        word_to_id = _load_vocab(vocab_path)
+
+    if not args.train_path:
+        train_path = os.path.join(data_path, "train")
+        train_data = _file_to_word_ids(listDir(train_path), word_to_id)
+    else:
+        train_path = args.train_path
+        train_data = _file_to_word_ids([train_path], word_to_id)
+    valid_path = os.path.join(data_path, "dev")
+    test_path = os.path.join(data_path, "dev")
+    valid_data = _file_to_word_ids(listDir(valid_path), word_to_id)
+    test_data = _file_to_word_ids(listDir(test_path), word_to_id)
+    vocabulary = len(word_to_id)
+    return train_data, valid_data, test_data, vocabulary
+
+
+def get_data_iter(raw_data, batch_size, num_steps):
+    def __impl__():
+        buf = []
+        while True:
+            if len(buf) >= num_steps * batch_size + 1:
+                x = np.asarray(
+                    buf[:-1], dtype='int64').reshape((batch_size, num_steps))
+                y = np.asarray(
+                    buf[1:], dtype='int64').reshape((batch_size, num_steps))
+                yield (x, y)
+                buf = [buf[-1]]
+            try:
+                buf.append(raw_data.next())
+            except StopIteration:
+                break
+
+    return __impl__

PaddleNLP/ELMO/run.sh

+export CUDA_VISIBLE_DEVICES=0 
+python  train.py \
+--train_path='baike/train/sentence_file_*'  \
+--test_path='baike/dev/sentence_file_*'  \
+--vocab_path baike/vocabulary_min5k.txt \
+--learning_rate 0.2 \
+--use_gpu True \
+--local True $@

PaddleNLP/ELMO/train.py

+#   Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+# 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 absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import six
+import numpy as np
+import time
+import os
+import math
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.core as core
+import paddle.fluid.framework as framework
+from paddle.fluid.executor import Executor
+import data
+from args import *
+import lm_model
+import logging
+logging.basicConfig()
+import pickle
+
+def prepare_batch_input(batch, args):
+    x = batch['token_ids']
+    x_r = batch['token_ids_reverse']
+    y = batch['next_token_id']
+    y_r = batch['next_token_id_reverse']
+    inst = []
+    for i in range(len(x)):
+        if args.use_custom_samples:
+            custom_samples_array = np.zeros(
+                (args.num_steps, args.n_negative_samples_batch + 1),
+                dtype='int64')
+            custom_samples_array_r = np.zeros(
+                (args.num_steps, args.n_negative_samples_batch + 1),
+                dtype='int64')
+            custom_probabilities_array = np.zeros(
+                (args.num_steps, args.n_negative_samples_batch + 1),
+                dtype='float32')
+            for j in range(args.num_steps):
+                for k in range(args.n_negative_samples_batch + 1):
+                    custom_samples_array[j][k] = k
+                    custom_samples_array_r[j][k] = k
+                    custom_probabilities_array[j][k] = 1.0
+                custom_samples_array[j][0] = y[i][j]
+                custom_samples_array_r[j][0] = y_r[i][j]
+            inst.append([
+                x[i], y[i], x_r[i], y_r[i], custom_samples_array,
+                custom_samples_array_r, custom_probabilities_array
+            ])
+        else:
+            inst.append([x[i], y[i], x_r[i], y_r[i]])
+    return inst
+
+
+def batch_reader(batch_list, args):
+    res = []
+    for batch in batch_list:
+        res.append(prepare_batch_input(batch, args))
+    return res
+
+
+def read_multiple(reader, batch_size, count, clip_last=True):
+    """
+    Stack data from reader for multi-devices.
+    """
+
+    def __impl__():
+        # one time read batch_size * count data for rnn
+        for data in reader():
+            inst_num_per_part = batch_size
+            split_data = {}
+            len_check = True
+            for k in data.keys():
+                if data[k] is not None:
+                    if len(data[k]) != batch_size * count:
+                        len_check = False
+                        print("data check error!!, data=" + data[k] +  ", k=" + k)
+                        break
+            if len_check:
+                res = []
+                for i in range(count):
+                    split_data = {}
+                    for k in data.keys():
+                        if data[k] is not None:
+                            split_data[k] = data[k][inst_num_per_part * i:inst_num_per_part * (i + 1)]
+                    res.append(split_data)
+                yield res
+	
+    return __impl__
+
+
+def LodTensor_Array(lod_tensor):
+    lod = lod_tensor.lod()
+    array = np.array(lod_tensor)
+    new_array = []
+    for i in range(len(lod[0]) - 1):
+        new_array.append(array[lod[0][i]:lod[0][i + 1]])
+    return new_array
+
+
+def get_current_model_para(train_prog, train_exe):
+    param_list = train_prog.block(0).all_parameters()
+    param_name_list = [p.name for p in param_list]
+
+    vals = {}
+    for p_name in param_name_list:
+        p_array = np.array(fluid.global_scope().find_var(p_name).get_tensor())
+        vals[p_name] = p_array
+
+    return vals
+
+
+def save_para_npz(train_prog, train_exe):
+    logger.info("begin to save model to model_base")
+    param_list = train_prog.block(0).all_parameters()
+    param_name_list = [p.name for p in param_list]
+
+    vals = {}
+    for p_name in param_name_list:
+        p_array = np.array(fluid.global_scope().find_var(p_name).get_tensor())
+        vals[p_name] = p_array
+
+    emb = vals["embedding_para"]
+    logger.info("begin to save model to model_base")
+    np.savez("mode_base", **vals)
+
+
+def prepare_input(batch, epoch_id=0, with_lr=True):
+    x, y = batch
+    inst = []
+    for i in range(len(x)):
+        inst.append([x[i], y[i]])
+    return inst
+
+
+def eval(vocab, infer_progs, dev_count, logger, args):
+    infer_prog, infer_startup_prog, infer_model = infer_progs
+    feed_order = infer_model.feed_order
+    loss = infer_model.loss
+
+    # prepare device
+    place = core.CUDAPlace(0) if args.use_gpu else core.CPUPlace()
+    exe = Executor(place)
+    if not args.use_gpu:
+        place = fluid.CPUPlace()
+        import multiprocessing
+        dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+    else:
+        place = fluid.CUDAPlace(0)
+        dev_count = fluid.core.get_cuda_device_count()
+
+    total_loss = 0.0
+    total_cnt = 0
+    n_batch_cnt = 0
+    n_batch_loss = 0.0
+    val_feed_list = [
+        infer_prog.global_block().var(var_name) for var_name in feed_order
+    ]
+    val_feeder = fluid.DataFeeder(val_feed_list, place)
+    dev_data = data.BidirectionalLMDataset(
+        args.test_path, vocab, test=True, shuffle_on_load=False)
+    dev_data_iter = lambda: dev_data.iter_batches(args.batch_size * dev_count, args.num_steps)
+    dev_reader = read_multiple(dev_data_iter, args.batch_size, dev_count)
+
+    last_hidden_values = np.zeros(
+        (dev_count, args.num_layers * 2 * args.batch_size * args.embed_size),
+        dtype='float32')
+    last_cell_values = np.zeros(
+        (dev_count, args.num_layers * 2 * args.batch_size * args.hidden_size),
+        dtype='float32')
+    for batch_id, batch_list in enumerate(dev_reader(), 1):
+        feed_data = batch_reader(batch_list, args)
+        feed = list(val_feeder.feed_parallel(feed_data, dev_count))
+        for i in range(dev_count):
+            init_hidden_tensor = fluid.core.LoDTensor()
+            if args.use_gpu:
+                placex = fluid.CUDAPlace(i)
+            else:
+                placex = fluid.CPUPlace()
+            init_hidden_tensor.set(last_hidden_values[i], placex)
+            init_cell_tensor = fluid.core.LoDTensor()
+            init_cell_tensor.set(last_cell_values[i], placex)
+
+            feed[i]['init_hiddens'] = init_hidden_tensor
+            feed[i]['init_cells'] = init_cell_tensor
+        last_hidden_values = []
+        last_cell_values = []
+        for i in range(dev_count):
+            val_fetch_outs = exe.run(
+                program=infer_prog,
+                feed=feed[i],
+                fetch_list=[
+                    infer_model.loss.name, infer_model.last_hidden.name,
+                    infer_model.last_cell.name
+                ],
+                return_numpy=False)
+            last_hidden_values.append(np.array(val_fetch_outs[1]))
+            last_cell_values.append(np.array(val_fetch_outs[2]))
+            total_loss += np.array(val_fetch_outs[0]).sum()
+
+            n_batch_cnt += len(np.array(val_fetch_outs[0]))
+            total_cnt += len(np.array(val_fetch_outs[0]))
+            n_batch_loss += np.array(val_fetch_outs[0]).sum()
+
+        last_hidden_values = np.array(last_hidden_values).reshape((
+            dev_count, args.num_layers * 2 * args.batch_size * args.embed_size))
+        last_cell_values = np.array(last_cell_values).reshape(
+            (dev_count,
+             args.num_layers * 2 * args.batch_size * args.hidden_size))
+
+        log_every_n_batch = args.log_interval
+        if log_every_n_batch > 0 and batch_id % log_every_n_batch == 0:
+            logger.info('Average dev loss from batch {} to {} is {}'.format(
+                batch_id - log_every_n_batch + 1, batch_id, "%.10f" % (
+                    n_batch_loss / n_batch_cnt)))
+            n_batch_loss = 0.0
+            n_batch_cnt = 0
+        batch_offset = 0
+
+    ppl = np.exp(total_loss / total_cnt)
+    return ppl
+
+
+def train():
+    args = parse_args()
+    if args.random_seed == 0:
+        args.random_seed = None
+        print("random seed is None")
+    if args.enable_ce:
+        random.seed(args.random_seed)
+        np.random.seed(args.random_seed)
+    logger = logging.getLogger("lm")
+    logger.setLevel(logging.INFO)
+    formatter = logging.Formatter(
+        '%(asctime)s - %(name)s - %(levelname)s - %(message)s')
+    console_handler = logging.StreamHandler()
+    console_handler.setLevel(logging.INFO)
+    console_handler.setFormatter(formatter)
+
+    logger.info('Running with args : {}'.format(args))
+    logger.info('Running paddle : {}'.format(paddle.version.commit))
+
+    hidden_size = args.hidden_size
+    batch_size = args.batch_size
+    data_path = args.data_path
+    logger.info("begin to load vocab")
+    vocab = data.Vocabulary(args.vocab_path, validate_file=True)
+    vocab_size = vocab.size
+    logger.info("finished load vocab")
+
+    logger.info('build the model...')
+    # build model
+    train_prog = fluid.Program()
+    train_startup_prog = fluid.Program()
+    if args.enable_ce:
+        train_prog.random_seed = args.random_seed
+        train_startup_prog.random_seed = args.random_seed
+
+    # build infer model
+    infer_prog = fluid.Program()
+    infer_startup_prog = fluid.Program()
+    with fluid.program_guard(infer_prog, infer_startup_prog):
+        with fluid.unique_name.guard():
+            # Infer process
+            infer_model = lm_model.LanguageModel(
+                args, vocab_size, test_mode=True)
+            infer_model.build()
+    infer_progs = infer_prog, infer_startup_prog, infer_model
+
+    with fluid.program_guard(train_prog, train_startup_prog):
+        with fluid.unique_name.guard():
+            # Training process
+            train_model = lm_model.LanguageModel(
+                args, vocab_size, test_mode=False)
+            train_model.build()
+            fluid.clip.set_gradient_clip(
+                clip=fluid.clip.GradientClipByGlobalNorm(
+                    clip_norm=args.max_grad_norm))
+
+            # build optimizer
+            if args.optim == 'adagrad':
+                optimizer = fluid.optimizer.Adagrad(
+                    learning_rate=args.learning_rate,
+                    epsilon=0.0,
+                    initial_accumulator_value=1.0)
+            elif args.optim == 'sgd':
+                optimizer = fluid.optimizer.SGD(
+                    learning_rate=args.learning_rate)
+            elif args.optim == 'adam':
+                optimizer = fluid.optimizer.Adam(
+                    learning_rate=args.learning_rate)
+            elif args.optim == 'rprop':
+                optimizer = fluid.optimizer.RMSPropOptimizer(
+                    learning_rate=args.learning_rate)
+            else:
+                logger.error('Unsupported optimizer: {}'.format(args.optim))
+                exit(-1)
+            optimizer.minimize(train_model.loss * args.num_steps)
+
+            # initialize parameters
+            place = core.CUDAPlace(0) if args.use_gpu else core.CPUPlace()
+            exe = Executor(place)
+    train_progs = train_prog, train_startup_prog, train_model
+
+    if args.local:
+        logger.info("local start_up:")
+        train_loop(args, logger, vocab, train_progs, infer_progs, optimizer)
+    else:
+        if args.update_method == "nccl2":
+            trainer_id = int(os.getenv("PADDLE_TRAINER_ID", "0"))
+            if args.test_nccl:
+                worker_endpoints_env = os.getenv("PADDLE_WORK_ENDPOINTS")
+                worker_endpoints = worker_endpoints_env.split(',')
+                trainers_num = len(worker_endpoints)
+                current_endpoint = worker_endpoints[trainer_id]
+            else:
+                port = os.getenv("PADDLE_PORT")
+                worker_ips = os.getenv("PADDLE_TRAINERS")
+                worker_endpoints = []
+                for ip in worker_ips.split(","):
+                    worker_endpoints.append(':'.join([ip, port]))
+                worker_endpoints_env = ','.join(worker_endpoints)
+                trainers_num = len(worker_endpoints)
+                current_endpoint = os.getenv("POD_IP") + ":" + port
+            if trainer_id == 0:
+                logger.info("train_id == 0, sleep 60s")
+                time.sleep(60)
+
+            logger.info("trainers_num:{}".format(trainers_num))
+            logger.info("worker_endpoints:{}".format(worker_endpoints))
+            logger.info("current_endpoint:{}".format(current_endpoint))
+            config = fluid.DistributeTranspilerConfig()
+            config.mode = "nccl2"
+            t = fluid.DistributeTranspiler(config=config)
+            t.transpile(
+                trainer_id,
+                trainers=worker_endpoints_env,
+                current_endpoint=current_endpoint,
+                program=train_prog,
+                startup_program=train_startup_prog)
+            train_progs = train_prog, train_startup_prog, train_model
+            train_loop(args, logger, vocab, train_progs, infer_progs, optimizer,
+                       trainers_num, trainer_id, worker_endpoints)
+        else:
+            port = os.getenv("PADDLE_PORT", "6174")
+            pserver_ips = os.getenv("PADDLE_PSERVERS") 
+            eplist = []
+            for ip in pserver_ips.split(","):
+                eplist.append(':'.join([ip, port]))
+            pserver_endpoints = ",".join(eplist) 
+            trainers = int(os.getenv("PADDLE_TRAINERS_NUM", "0"))
+            current_endpoint = os.getenv("POD_IP") + ":" + port
+            trainer_id = int(os.getenv("PADDLE_TRAINER_ID"))
+
+            logger.info("pserver_endpoints:{}".format(pserver_endpoints))
+            logger.info("current_endpoint:{}".format(current_endpoint))
+            logger.info("trainer_id:{}".format(trainer_id))
+            logger.info("pserver_ips:{}".format(pserver_ips))
+            logger.info("port:{}".format(port))
+
+            t = fluid.DistributeTranspiler()
+            t.transpile(
+                trainer_id,
+                pservers=pserver_endpoints,
+                trainers=trainers,
+                program=train_prog,
+                startup_program=startup_prog)
+
+            if training_role == "PSERVER":
+                logger.info("distributed: pserver started")
+                current_endpoint = os.getenv("POD_IP") + ":" + os.getenv(
+                    "PADDLE_PORT")
+                if not current_endpoint:
+                    logger.critical("need env SERVER_ENDPOINT")
+                    exit(1)
+                pserver_prog = t.get_pserver_program(current_endpoint)
+                pserver_startup = t.get_startup_program(current_endpoint,
+                                                        pserver_prog)
+
+                exe.run(pserver_startup)
+                exe.run(pserver_prog)
+            elif training_role == "TRAINER":
+                logger.info("distributed: trainer started")
+                trainer_prog = t.get_trainer_program()
+                train_loop(args, logger, vocab, train_progs, infer_progs,
+                           optimizer)
+            else:
+                logger.critical(
+                    "environment var TRAINER_ROLE should be TRAINER os PSERVER")
+                exit(1)
+
+def init_pretraining_params(exe,
+                            pretraining_params_path,
+                            main_program):
+    assert os.path.exists(pretraining_params_path
+                          ), "[%s] cann't be found." % pretraining_params_path
+
+    def existed_params(var):
+        if not isinstance(var, fluid.framework.Parameter):
+            return False
+        return os.path.exists(os.path.join(pretraining_params_path, var.name))
+
+    fluid.io.load_vars(
+        exe,
+        pretraining_params_path,
+        main_program=main_program,
+        predicate=existed_params)
+    print("Load pretraining parameters from {}.".format(
+        pretraining_params_path))
+
+
+def train_loop(args,
+               logger,
+               vocab,
+               train_progs,
+               infer_progs,
+               optimizer,
+               nccl2_num_trainers=1,
+               nccl2_trainer_id=0,
+               worker_endpoints=None):
+    train_prog, train_startup_prog, train_model = train_progs
+    infer_prog, infer_startup_prog, infer_model = infer_progs
+
+    # prepare device
+    place = core.CUDAPlace(0) if args.use_gpu else core.CPUPlace()
+    exe = Executor(place)
+    if not args.use_gpu:
+        place = fluid.CPUPlace()
+        import multiprocessing
+        dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+    else:
+        place = fluid.CUDAPlace(0)
+        dev_count = fluid.core.get_cuda_device_count()
+
+    if args.load_dir:
+        logger.info('load pretrained checkpoints from {}'.format(args.load_dir))
+        fluid.io.load_persistables(exe, args.load_dir, main_program=train_prog)
+    elif args.load_pretraining_params:
+        logger.info('load pretrained params from {}'.format(args.load_pretraining_params))
+        exe.run(train_startup_prog)
+        init_pretraining_params(exe, args.load_pretraining_params, main_program=train_prog)
+    else:
+        exe.run(train_startup_prog)
+
+    # prepare data
+    feed_list = [
+        train_prog.global_block().var(var_name)
+        for var_name in train_model.feed_order
+    ]
+    feeder = fluid.DataFeeder(feed_list, place)
+
+    logger.info('Training the model...')
+    exe_strategy = fluid.parallel_executor.ExecutionStrategy()
+    parallel_executor = fluid.ParallelExecutor(
+        loss_name=train_model.loss.name,
+        main_program=train_prog,
+        use_cuda=bool(args.use_gpu),
+        exec_strategy=exe_strategy,
+        num_trainers=nccl2_num_trainers,
+        trainer_id=nccl2_trainer_id)
+
+    logger.info("begin to load data")
+    train_data = data.BidirectionalLMDataset(
+        args.train_path,
+        vocab,
+        test=(not args.shuffle),
+        shuffle_on_load=args.shuffle)
+    logger.info("finished load vocab")
+
+    # get train epoch size
+    log_interval = args.log_interval
+    total_time = 0.0
+    batch_size = args.batch_size
+    hidden_size = args.hidden_size
+    custom_samples_array = np.zeros(
+        (batch_size, args.num_steps, args.n_negative_samples_batch + 1),
+        dtype='int64')
+    custom_probabilities_array = np.zeros(
+        (batch_size, args.num_steps, args.n_negative_samples_batch + 1),
+        dtype='float32')
+    for i in range(batch_size):
+        for j in range(0, args.num_steps):
+            for k in range(0, args.n_negative_samples_batch + 1):
+                custom_samples_array[i][j][k] = k
+                custom_probabilities_array[i][j][k] = 1.0
+
+    for epoch_id in range(args.max_epoch):
+        start_time = time.time()
+        logger.info("epoch id {}".format(epoch_id))
+        train_data_iter = lambda: train_data.iter_batches(batch_size * dev_count, args.num_steps)
+        train_reader = read_multiple(train_data_iter, batch_size, dev_count)
+
+        total_num = 0
+        n_batch_loss = 0.0
+        n_batch_cnt = 0
+        last_hidden_values = np.zeros(
+            (dev_count, args.num_layers * 2 * batch_size * args.embed_size),
+            dtype='float32')
+        last_cell_values = np.zeros(
+            (dev_count, args.num_layers * 2 * batch_size * hidden_size),
+            dtype='float32')
+
+        begin_time = time.time()
+        for batch_id, batch_list in enumerate(train_reader(), 1):
+            feed_data = batch_reader(batch_list, args)
+            feed = list(feeder.feed_parallel(feed_data, dev_count))
+            for i in range(dev_count):
+                init_hidden_tensor = fluid.core.LoDTensor()
+                if args.use_gpu:
+                    placex = fluid.CUDAPlace(i)
+                else:
+                    placex = fluid.CPUPlace()
+                init_hidden_tensor.set(last_hidden_values[i], placex)
+                init_cell_tensor = fluid.core.LoDTensor()
+                init_cell_tensor.set(last_cell_values[i], placex)
+
+                feed[i]['init_hiddens'] = init_hidden_tensor
+                feed[i]['init_cells'] = init_cell_tensor
+
+            fetch_outs = parallel_executor.run(
+                feed=feed,
+                fetch_list=[
+                    train_model.loss.name, train_model.last_hidden.name,
+                    train_model.last_cell.name
+                ], 
+                return_numpy=False)
+            cost_train = np.array(fetch_outs[0]).mean()
+            last_hidden_values = np.array(fetch_outs[1])
+            last_hidden_values = last_hidden_values.reshape(
+                (dev_count, args.num_layers * 2 * batch_size * args.embed_size))
+            last_cell_values = np.array(fetch_outs[2])
+            last_cell_values = last_cell_values.reshape((
+                dev_count, args.num_layers * 2 * batch_size * args.hidden_size))
+
+            total_num += args.batch_size * dev_count
+            n_batch_loss += np.array(fetch_outs[0]).sum()
+
+            n_batch_cnt += len(np.array(fetch_outs[0]))
+
+            if batch_id > 0 and batch_id % log_interval == 0:
+                smoothed_ppl = np.exp(n_batch_loss / n_batch_cnt)
+                ppl = np.exp(
+                    np.array(fetch_outs[0]).sum() /
+                    len(np.array(fetch_outs[0])))
+                used_time = time.time() - begin_time
+                speed = log_interval / used_time
+                logger.info(
+                    "[train] epoch:{}, step:{}, loss:{:.3f}, ppl:{:.3f}, smoothed_ppl:{:.3f}, speed:{:.3f}".
+                    format(epoch_id, batch_id, n_batch_loss / n_batch_cnt, ppl,
+                           smoothed_ppl, speed))
+                n_batch_loss = 0.0
+                n_batch_cnt = 0
+                begin_time = time.time()
+            if batch_id > 0 and batch_id % args.dev_interval == 0:
+                valid_ppl = eval(vocab, infer_progs, dev_count, logger, args)
+                logger.info("valid ppl {}".format(valid_ppl))
+            if batch_id > 0 and batch_id % args.save_interval == 0:
+                model_path = os.path.join(args.para_save_dir,
+                                          str(batch_id + epoch_id))
+                if not os.path.isdir(model_path):
+                    os.makedirs(model_path)
+                fluid.io.save_persistables(
+                    executor=exe, dirname=model_path, main_program=train_prog)
+
+        end_time = time.time()
+        total_time += end_time - start_time
+        logger.info("train ppl {}".format(ppl))
+
+        if epoch_id == args.max_epoch - 1 and args.enable_ce:
+            logger.info("lstm_language_model_duration\t%s" %
+                        (total_time / args.max_epoch))
+            logger.info("lstm_language_model_loss\t%s" % ppl[0])
+
+        model_path = os.path.join(args.para_save_dir, str(epoch_id))
+        if not os.path.isdir(model_path):
+            os.makedirs(model_path)
+        fluid.io.save_persistables(
+            executor=exe, dirname=model_path, main_program=train_prog)
+        valid_ppl = eval(vocab, infer_progs, dev_count, logger, args)
+        logger.info("valid ppl {}".format(valid_ppl))
+    test_ppl = eval(vocab, infer_progs, dev_count, logger, args)
+    logger.info("test ppl {}".format(test_ppl))
+
+
+if __name__ == '__main__':
+    train()