restructure the code of scheduled sampling

scheduled_sampling/README.md

@@ -37,7 +37,7 @@ Scheduled Sampling主要应用在序列到序列模型的训练阶段，而生
 
 ## 模型实现
 
-由于Scheduled Sampling是对序列到序列模型的改进，其整体实现框架与序列到序列模型较为相似。为突出本文重点，这里仅介绍与Scheduled Sampling相关的部分，完整的代码见`scheduled_sampling.py`。
+由于Scheduled Sampling是对序列到序列模型的改进，其整体实现框架与序列到序列模型较为相似。为突出本文重点，这里仅介绍与Scheduled Sampling相关的部分，完整的代码见`network_conf.py`。
 
 首先导入需要的包，并定义控制衰减概率的类`RandomScheduleGenerator`，如下：
 
@@ -119,9 +119,10 @@ true_token_flags = paddle.layer.data(
 这里还需要对原始reader进行封装，增加`true_token_flag`的数据生成器。下面以线性衰减为例说明如何调用上面定义的`RandomScheduleGenerator`产生`true_token_flag`的输入数据。
 
 ```python
-schedule_generator = RandomScheduleGenerator("linear", 0.75, 1000000)
-
-def gen_schedule_data(reader):
+def gen_schedule_data(reader,
+                      schedule_type="linear",
+                      decay_a=0.75,
+                      decay_b=1000000):
     """
     Creates a data reader for scheduled sampling.
 
@@ -130,10 +131,17 @@ def gen_schedule_data(reader):
 
     :param reader: the original reader.
     :type reader: callable
+    :param schedule_type: the type of sampling rate decay.
+    :type schedule_type: str
+    :param decay_a: the decay parameter a.
+    :type decay_a: float
+    :param decay_b: the decay parameter b.
+    :type decay_b: float
 
     :return: the new reader with the field "true_token_flag".
     :rtype: callable
     """
+    schedule_generator = RandomScheduleGenerator(schedule_type, decay_a, decay_b)
 
     def data_reader():
         for src_ids, trg_ids, trg_ids_next in reader():
@@ -149,61 +157,60 @@ def gen_schedule_data(reader):
 
 ```python
 def gru_decoder_with_attention_train(enc_vec, enc_proj, true_word,
-                                     true_token_flag):
-    """
-    The decoder step for training.
-    :param enc_vec: the encoder vector for attention
-    :type enc_vec: LayerOutput
-    :param enc_proj: the encoder projection for attention
-    :type enc_proj: LayerOutput
-    :param true_word: the ground-truth target word
-    :type true_word: LayerOutput
-    :param true_token_flag: the flag of using the ground-truth target word
-    :type true_token_flag: LayerOutput
-    :return: the softmax output layer
-    :rtype: LayerOutput
-    """
-
-    decoder_mem = paddle.layer.memory(
-        name='gru_decoder', size=decoder_size, boot_layer=decoder_boot)
-
-    context = paddle.networks.simple_attention(
-        encoded_sequence=enc_vec,
-        encoded_proj=enc_proj,
-        decoder_state=decoder_mem)
-
-    gru_out_memory = paddle.layer.memory(
-        name='gru_out', size=target_dict_dim)
-
-    generated_word = paddle.layer.max_id(input=gru_out_memory)
-
-    generated_word_emb = paddle.layer.embedding(
-        input=generated_word,
-        size=word_vector_dim,
-        param_attr=paddle.attr.ParamAttr(name='_target_language_embedding'))
-
-    current_word = paddle.layer.multiplex(
-        input=[true_token_flag, true_word, generated_word_emb])
-
-    with paddle.layer.mixed(size=decoder_size * 3) as decoder_inputs:
-        decoder_inputs += paddle.layer.full_matrix_projection(input=context)
-        decoder_inputs += paddle.layer.full_matrix_projection(
-            input=current_word)
-
-    gru_step = paddle.layer.gru_step(
-        name='gru_decoder',
-        input=decoder_inputs,
-        output_mem=decoder_mem,
-        size=decoder_size)
-
-    with paddle.layer.mixed(
-            name='gru_out',
-            size=target_dict_dim,
-            bias_attr=True,
-            act=paddle.activation.Softmax()) as out:
-        out += paddle.layer.full_matrix_projection(input=gru_step)
-
-    return out
+                                       true_token_flag):
+      """
+      The decoder step for training.
+      :param enc_vec: the encoder vector for attention
+      :type enc_vec: LayerOutput
+      :param enc_proj: the encoder projection for attention
+      :type enc_proj: LayerOutput
+      :param true_word: the ground-truth target word
+      :type true_word: LayerOutput
+      :param true_token_flag: the flag of using the ground-truth target word
+      :type true_token_flag: LayerOutput
+      :return: the softmax output layer
+      :rtype: LayerOutput
+      """
+
+      decoder_mem = paddle.layer.memory(
+          name='gru_decoder', size=decoder_size, boot_layer=decoder_boot)
+
+      context = paddle.networks.simple_attention(
+          encoded_sequence=enc_vec,
+          encoded_proj=enc_proj,
+          decoder_state=decoder_mem)
+
+      gru_out_memory = paddle.layer.memory(
+          name='gru_out', size=target_dict_dim)
+
+      generated_word = paddle.layer.max_id(input=gru_out_memory)
+
+      generated_word_emb = paddle.layer.embedding(
+          input=generated_word,
+          size=word_vector_dim,
+          param_attr=paddle.attr.ParamAttr(name='_target_language_embedding'))
+
+      current_word = paddle.layer.multiplex(
+          input=[true_token_flag, true_word, generated_word_emb])
+
+      decoder_inputs = paddle.layer.fc(
+          input=[context, current_word],
+          size=decoder_size * 3,
+          act=paddle.activation.Linear(),
+          bias_attr=False)
+
+      gru_step = paddle.layer.gru_step(
+          name='gru_decoder',
+          input=decoder_inputs,
+          output_mem=decoder_mem,
+          size=decoder_size)
+
+      out = paddle.layer.fc(
+          name='gru_out',
+          input=gru_step,
+          size=target_dict_dim,
+          act=paddle.activation.Softmax())
+      return out
 ```
 
 该函数使用`memory`层`gru_out_memory`记忆上一时刻生成的元素，根据`gru_out_memory`选择概率最大的词语`generated_word`作为生成的词语。`multiplex`层会在真实元素`true_word`和生成的元素`generated_word`之间做出选择，并将选择的结果作为解码器输入。`multiplex`层使用了三个输入，分别为`true_token_flag`、`true_word`和`generated_word_emb`。对于这三个输入中每个元素，若`true_token_flag`中的值为`0`，则`multiplex`层输出`true_word`中的相应元素；若`true_token_flag`中的值为`1`，则`multiplex`层输出`generated_word_emb`中的相应元素。

scheduled_sampling/generate.py

+import gzip
+import argparse
+import distutils.util
+import paddle.v2 as paddle
+
+from network_conf import seqToseq_net
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description="PaddlePaddle Scheduled Sampling")
+    parser.add_argument(
+        '--model_path',
+        type=str,
+        required=True,
+        help="The path for trained model to load.")
+    parser.add_argument(
+        '--beam_size',
+        type=int,
+        default=3,
+        help='The width of beam expansion. (default: %(default)s)')
+    parser.add_argument(
+        "--use_gpu",
+        type=distutils.util.strtobool,
+        default=False,
+        help="Use gpu or not. (default: %(default)s)")
+    parser.add_argument(
+        "--trainer_count",
+        type=int,
+        default=1,
+        help="Trainer number. (default: %(default)s)")
+
+    return parser.parse_args()
+
+
+def generate(gen_data, dict_size, model_path, beam_size):
+    beam_gen = seqToseq_net(dict_size, dict_size, beam_size, is_generating=True)
+
+    with gzip.open(model_path, 'r') as f:
+        parameters = paddle.parameters.Parameters.from_tar(f)
+
+    # prob is the prediction probabilities, and id is the prediction word.
+    beam_result = paddle.infer(
+        output_layer=beam_gen,
+        parameters=parameters,
+        input=gen_data,
+        field=['prob', 'id'])
+
+    # get the dictionary
+    src_dict, trg_dict = paddle.dataset.wmt14.get_dict(dict_size)
+
+    # the delimited element of generated sequences is -1,
+    # the first element of each generated sequence is the sequence length
+    seq_list = []
+    seq = []
+    for w in beam_result[1]:
+        if w != -1:
+            seq.append(w)
+        else:
+            seq_list.append(' '.join([trg_dict.get(w) for w in seq[1:]]))
+            seq = []
+
+    prob = beam_result[0]
+    for i in xrange(gen_num):
+        print "\n*******************************************************\n"
+        print "src:", ' '.join([src_dict.get(w) for w in gen_data[i][0]]), "\n"
+        for j in xrange(beam_size):
+            print "prob = %f:" % (prob[i][j]), seq_list[i * beam_size + j]
+
+
+if __name__ == '__main__':
+    args = parse_args()
+
+    dict_size = 30000
+
+    paddle.init(use_gpu=args.use_gpu, trainer_count=args.trainer_count)
+
+    # use the first 3 samples for generation
+    gen_creator = paddle.dataset.wmt14.gen(dict_size)
+    gen_data = []
+    gen_num = 3
+    for item in gen_creator():
+        gen_data.append((item[0], ))
+        if len(gen_data) == gen_num:
+            break
+
+    generate(
+        gen_data,
+        dict_size=dict_size,
+        model_path=args.model_path,
+        beam_size=args.beam_size)

scheduled_sampling/scheduled_sampling.py → scheduled_sampling/network_conf.py

-import sys
 import paddle.v2 as paddle
-from random_schedule_generator import RandomScheduleGenerator
 
-schedule_generator = RandomScheduleGenerator("linear", 0.75, 1000000)
+__all__ = ["seqToseq_net"]
 
+### Network Architecture
+word_vector_dim = 512  # dimension of word vector
+decoder_size = 512  # dimension of hidden unit in GRU Decoder network
+encoder_size = 512  # dimension of hidden unit in GRU Encoder network
 
-def gen_schedule_data(reader):
-    """
-    Creates a data reader for scheduled sampling.
-
-    Output from the iterator that created by original reader will be
-    appended with "true_token_flag" to indicate whether to use true token.
+max_length = 250
 
-    :param reader: the original reader.
-    :type reader: callable
 
-    :return: the new reader with the field "true_token_flag".
-    :rtype: callable
-    """
-
-    def data_reader():
-        for src_ids, trg_ids, trg_ids_next in reader():
-            yield src_ids, trg_ids, trg_ids_next, \
-                  [0] + schedule_generator.processBatch(len(trg_ids) - 1)
-
-    return data_reader
-
-
-def seqToseq_net(source_dict_dim, target_dict_dim, is_generating=False):
+def seqToseq_net(source_dict_dim,
+                 target_dict_dim,
+                 beam_size,
+                 is_generating=False):
     """
     The definition of the sequence to sequence model
     :param source_dict_dim: the dictionary size of the source language
     :type source_dict_dim: int
     :param target_dict_dim: the dictionary size of the target language
     :type target_dict_dim: int
+    :param beam_size: The width of beam expansion
+    :type beam_size: int
     :param is_generating: whether in generating mode
     :type is_generating: Bool
     :return: the last layer of the network
     :rtype: LayerOutput
     """
-    ### Network Architecture
-    word_vector_dim = 512  # dimension of word vector
-    decoder_size = 512  # dimension of hidden unit in GRU Decoder network
-    encoder_size = 512  # dimension of hidden unit in GRU Encoder network
-
-    beam_size = 3
-    max_length = 250
 
     #### Encoder
     src_word_id = paddle.layer.data(
@@ -55,21 +36,24 @@ def seqToseq_net(source_dict_dim, target_dict_dim, is_generating=False):
         input=src_word_id, size=word_vector_dim)
     src_forward = paddle.networks.simple_gru(
         input=src_embedding, size=encoder_size)
-    src_backward = paddle.networks.simple_gru(
+    src_reverse = paddle.networks.simple_gru(
         input=src_embedding, size=encoder_size, reverse=True)
-    encoded_vector = paddle.layer.concat(input=[src_forward, src_backward])
+    encoded_vector = paddle.layer.concat(input=[src_forward, src_reverse])
 
     #### Decoder
-    with paddle.layer.mixed(size=decoder_size) as encoded_proj:
-        encoded_proj += paddle.layer.full_matrix_projection(
-            input=encoded_vector)
+    encoded_proj = paddle.layer.fc(
+        input=encoded_vector,
+        size=decoder_size,
+        act=paddle.activation.Linear(),
+        bias_attr=False)
 
-    backward_first = paddle.layer.first_seq(input=src_backward)
+    reverse_first = paddle.layer.first_seq(input=src_reverse)
 
-    with paddle.layer.mixed(
-            size=decoder_size, act=paddle.activation.Tanh()) as decoder_boot:
-        decoder_boot += paddle.layer.full_matrix_projection(
-            input=backward_first)
+    decoder_boot = paddle.layer.fc(
+        input=reverse_first,
+        size=decoder_size,
+        act=paddle.activation.Tanh(),
+        bias_attr=False)
 
     def gru_decoder_with_attention_train(enc_vec, enc_proj, true_word,
                                          true_token_flag):
@@ -108,10 +92,11 @@ def seqToseq_net(source_dict_dim, target_dict_dim, is_generating=False):
         current_word = paddle.layer.multiplex(
             input=[true_token_flag, true_word, generated_word_emb])
 
-        with paddle.layer.mixed(size=decoder_size * 3) as decoder_inputs:
-            decoder_inputs += paddle.layer.full_matrix_projection(input=context)
-            decoder_inputs += paddle.layer.full_matrix_projection(
-                input=current_word)
+        decoder_inputs = paddle.layer.fc(
+            input=[context, current_word],
+            size=decoder_size * 3,
+            act=paddle.activation.Linear(),
+            bias_attr=False)
 
         gru_step = paddle.layer.gru_step(
             name='gru_decoder',
@@ -119,16 +104,14 @@ def seqToseq_net(source_dict_dim, target_dict_dim, is_generating=False):
             output_mem=decoder_mem,
             size=decoder_size)
 
-        with paddle.layer.mixed(
-                name='gru_out',
-                size=target_dict_dim,
-                bias_attr=True,
-                act=paddle.activation.Softmax()) as out:
-            out += paddle.layer.full_matrix_projection(input=gru_step)
-
+        out = paddle.layer.fc(
+            name='gru_out',
+            input=gru_step,
+            size=target_dict_dim,
+            act=paddle.activation.Softmax())
         return out
 
-    def gru_decoder_with_attention_test(enc_vec, enc_proj, current_word):
+    def gru_decoder_with_attention_gen(enc_vec, enc_proj, current_word):
         """
         The decoder step for generating.
         :param enc_vec: the encoder vector for attention
@@ -149,10 +132,11 @@ def seqToseq_net(source_dict_dim, target_dict_dim, is_generating=False):
             encoded_proj=enc_proj,
             decoder_state=decoder_mem)
 
-        with paddle.layer.mixed(size=decoder_size * 3) as decoder_inputs:
-            decoder_inputs += paddle.layer.full_matrix_projection(input=context)
-            decoder_inputs += paddle.layer.full_matrix_projection(
-                input=current_word)
+        decoder_inputs = paddle.layer.fc(
+            input=[context, current_word],
+            size=decoder_size * 3,
+            act=paddle.activation.Linear(),
+            bias_attr=False)
 
         gru_step = paddle.layer.gru_step(
             name='gru_decoder',
@@ -160,17 +144,16 @@ def seqToseq_net(source_dict_dim, target_dict_dim, is_generating=False):
             output_mem=decoder_mem,
             size=decoder_size)
 
-        with paddle.layer.mixed(
-                size=target_dict_dim,
-                bias_attr=True,
-                act=paddle.activation.Softmax()) as out:
-            out += paddle.layer.full_matrix_projection(input=gru_step)
+        out = paddle.layer.fc(
+            name='gru_out',
+            input=gru_step,
+            size=target_dict_dim,
+            act=paddle.activation.Softmax())
         return out
 
     decoder_group_name = "decoder_group"
     group_input1 = paddle.layer.StaticInput(input=encoded_vector, is_seq=True)
     group_input2 = paddle.layer.StaticInput(input=encoded_proj, is_seq=True)
-    group_inputs = [group_input1, group_input2]
 
     if not is_generating:
         trg_embedding = paddle.layer.embedding(
@@ -179,12 +162,14 @@ def seqToseq_net(source_dict_dim, target_dict_dim, is_generating=False):
                 type=paddle.data_type.integer_value_sequence(target_dict_dim)),
             size=word_vector_dim,
             param_attr=paddle.attr.ParamAttr(name='_target_language_embedding'))
-        group_inputs.append(trg_embedding)
 
         true_token_flags = paddle.layer.data(
             name='true_token_flag',
             type=paddle.data_type.integer_value_sequence(2))
-        group_inputs.append(true_token_flags)
+
+        group_inputs = [
+            group_input1, group_input2, trg_embedding, true_token_flags
+        ]
 
         decoder = paddle.layer.recurrent_group(
             name=decoder_group_name,
@@ -194,6 +179,7 @@ def seqToseq_net(source_dict_dim, target_dict_dim, is_generating=False):
         lbl = paddle.layer.data(
             name='target_language_next_word',
             type=paddle.data_type.integer_value_sequence(target_dict_dim))
+
         cost = paddle.layer.classification_cost(input=decoder, label=lbl)
 
         return cost
@@ -202,122 +188,15 @@ def seqToseq_net(source_dict_dim, target_dict_dim, is_generating=False):
             size=target_dict_dim,
             embedding_name='_target_language_embedding',
             embedding_size=word_vector_dim)
-        group_inputs.append(trg_embedding)
+
+        group_inputs = [group_input1, group_input2, trg_embedding]
 
         beam_gen = paddle.layer.beam_search(
             name=decoder_group_name,
-            step=gru_decoder_with_attention_test,
+            step=gru_decoder_with_attention_gen,
             input=group_inputs,
             bos_id=0,
             eos_id=1,
             beam_size=beam_size,
             max_length=max_length)
-
         return beam_gen
-
-
-def main():
-    paddle.init(use_gpu=False, trainer_count=1)
-    is_generating = False
-    model_path_for_generating = 'params_pass_1.tar.gz'
-
-    # source and target dict dim.
-    dict_size = 30000
-    source_dict_dim = target_dict_dim = dict_size
-
-    # train the network
-    if not is_generating:
-        cost = seqToseq_net(source_dict_dim, target_dict_dim)
-        parameters = paddle.parameters.create(cost)
-
-        # define optimize method and trainer
-        optimizer = paddle.optimizer.Adam(
-            learning_rate=5e-5,
-            regularization=paddle.optimizer.L2Regularization(rate=8e-4))
-        trainer = paddle.trainer.SGD(
-            cost=cost, parameters=parameters, update_equation=optimizer)
-        # define data reader
-        wmt14_reader = paddle.batch(
-            gen_schedule_data(
-                paddle.reader.shuffle(
-                    paddle.dataset.wmt14.train(dict_size), buf_size=8192)),
-            batch_size=5)
-
-        feeding = {
-            'source_language_word': 0,
-            'target_language_word': 1,
-            'target_language_next_word': 2,
-            'true_token_flag': 3
-        }
-
-        # define event_handler callback
-        def event_handler(event):
-            if isinstance(event, paddle.event.EndIteration):
-                if event.batch_id % 10 == 0:
-                    print "\nPass %d, Batch %d, Cost %f, %s" % (
-                        event.pass_id, event.batch_id, event.cost,
-                        event.metrics)
-                else:
-                    sys.stdout.write('.')
-                    sys.stdout.flush()
-            if isinstance(event, paddle.event.EndPass):
-                # save parameters
-                with gzip.open('params_pass_%d.tar.gz' % event.pass_id,
-                               'w') as f:
-                    trainer.save_parameter_to_tar(f)
-
-        # start to train
-        trainer.train(
-            reader=wmt14_reader,
-            event_handler=event_handler,
-            feeding=feeding,
-            num_passes=2)
-
-    # generate a english sequence to french
-    else:
-        # use the first 3 samples for generation
-        gen_creator = paddle.dataset.wmt14.gen(dict_size)
-        gen_data = []
-        gen_num = 3
-        for item in gen_creator():
-            gen_data.append((item[0], ))
-            if len(gen_data) == gen_num:
-                break
-
-        beam_gen = seqToseq_net(source_dict_dim, target_dict_dim, is_generating)
-        # get the trained model
-        with gzip.open(model_path_for_generating, 'r') as f:
-            parameters = Parameters.from_tar(f)
-        # prob is the prediction probabilities, and id is the prediction word.
-        beam_result = paddle.infer(
-            output_layer=beam_gen,
-            parameters=parameters,
-            input=gen_data,
-            field=['prob', 'id'])
-
-        # get the dictionary
-        src_dict, trg_dict = paddle.dataset.wmt14.get_dict(dict_size)
-
-        # the delimited element of generated sequences is -1,
-        # the first element of each generated sequence is the sequence length
-        seq_list = []
-        seq = []
-        for w in beam_result[1]:
-            if w != -1:
-                seq.append(w)
-            else:
-                seq_list.append(' '.join([trg_dict.get(w) for w in seq[1:]]))
-                seq = []
-
-        prob = beam_result[0]
-        beam_size = 3
-        for i in xrange(gen_num):
-            print "\n*******************************************************\n"
-            print "src:", ' '.join(
-                [src_dict.get(w) for w in gen_data[i][0]]), "\n"
-            for j in xrange(beam_size):
-                print "prob = %f:" % (prob[i][j]), seq_list[i * beam_size + j]
-
-
-if __name__ == '__main__':
-    main()

scheduled_sampling/reader.py

+from utils import RandomScheduleGenerator
+
+
+def gen_schedule_data(reader,
+                      schedule_type="linear",
+                      decay_a=0.75,
+                      decay_b=1000000):
+    """
+    Creates a data reader for scheduled sampling.
+
+    Output from the iterator that created by original reader will be
+    appended with "true_token_flag" to indicate whether to use true token.
+
+    :param reader: the original reader.
+    :type reader: callable
+    :param schedule_type: the type of sampling rate decay.
+    :type schedule_type: str
+    :param decay_a: the decay parameter a.
+    :type decay_a: float
+    :param decay_b: the decay parameter b.
+    :type decay_b: float
+
+    :return: the new reader with the field "true_token_flag".
+    :rtype: callable
+    """
+    schedule_generator = RandomScheduleGenerator(schedule_type, decay_a,
+                                                 decay_b)
+
+    def data_reader():
+        for src_ids, trg_ids, trg_ids_next in reader():
+            yield src_ids, trg_ids, trg_ids_next, \
+                  [0] + schedule_generator.processBatch(len(trg_ids) - 1)
+
+    return data_reader
+
+
+feeding = {
+    'source_language_word': 0,
+    'target_language_word': 1,
+    'target_language_next_word': 2,
+    'true_token_flag': 3
+}

scheduled_sampling/train.py

+import os
+import sys
+import gzip
+import argparse
+import distutils.util
+import paddle.v2 as paddle
+
+import reader
+from network_conf import seqToseq_net
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description="PaddlePaddle Scheduled Sampling")
+    parser.add_argument(
+        '--schedule_type',
+        type=str,
+        default="linear",
+        help='The type of sampling rate decay. Supported type: constant, linear, exponential, inverse_sigmoid. (default: %(default)s)'
+    )
+    parser.add_argument(
+        '--decay_a',
+        type=float,
+        default=0.75,
+        help='The sampling rate decay parameter a. (default: %(default)s)')
+    parser.add_argument(
+        '--decay_b',
+        type=float,
+        default=1000000,
+        help='The sampling rate decay parameter b. (default: %(default)s)')
+    parser.add_argument(
+        '--beam_size',
+        type=int,
+        default=3,
+        help='The width of beam expansion. (default: %(default)s)')
+    parser.add_argument(
+        "--use_gpu",
+        type=distutils.util.strtobool,
+        default=False,
+        help="Use gpu or not. (default: %(default)s)")
+    parser.add_argument(
+        "--trainer_count",
+        type=int,
+        default=1,
+        help="Trainer number. (default: %(default)s)")
+    parser.add_argument(
+        '--batch_size',
+        type=int,
+        default=32,
+        help="Size of a mini-batch. (default: %(default)s)")
+    parser.add_argument(
+        '--num_passes',
+        type=int,
+        default=10,
+        help="Number of passes to train. (default: %(default)s)")
+    parser.add_argument(
+        '--model_output_dir',
+        type=str,
+        default='models',
+        help="The path for model to store. (default: %(default)s)")
+
+    return parser.parse_args()
+
+
+def train(dict_size, batch_size, num_passes, beam_size, schedule_type, decay_a,
+          decay_b, model_dir):
+    optimizer = paddle.optimizer.Adam(
+        learning_rate=1e-4,
+        regularization=paddle.optimizer.L2Regularization(rate=1e-5))
+
+    cost = seqToseq_net(dict_size, dict_size, beam_size)
+
+    parameters = paddle.parameters.create(cost)
+
+    trainer = paddle.trainer.SGD(
+        cost=cost, parameters=parameters, update_equation=optimizer)
+
+    wmt14_reader = reader.gen_schedule_data(
+        paddle.reader.shuffle(
+            paddle.dataset.wmt14.train(dict_size), buf_size=8192),
+        schedule_type, decay_a, decay_b)
+
+    # define event_handler callback
+    def event_handler(event):
+        if isinstance(event, paddle.event.EndIteration):
+            if event.batch_id % 10 == 0:
+                print "\nPass %d, Batch %d, Cost %f, %s" % (
+                    event.pass_id, event.batch_id, event.cost, event.metrics)
+            else:
+                sys.stdout.write('.')
+                sys.stdout.flush()
+        if isinstance(event, paddle.event.EndPass):
+            # save parameters
+            with gzip.open(
+                    os.path.join(model_dir, 'params_pass_%d.tar.gz' %
+                                 event.pass_id), 'w') as f:
+                trainer.save_parameter_to_tar(f)
+
+    # start to train
+    trainer.train(
+        reader=paddle.batch(wmt14_reader, batch_size=batch_size),
+        event_handler=event_handler,
+        feeding=reader.feeding,
+        num_passes=num_passes)
+
+
+if __name__ == '__main__':
+    args = parse_args()
+
+    if not os.path.isdir(args.model_output_dir):
+        os.mkdir(args.model_output_dir)
+
+    paddle.init(use_gpu=args.use_gpu, trainer_count=args.trainer_count)
+
+    train(
+        dict_size=30000,
+        batch_size=args.batch_size,
+        num_passes=args.num_passes,
+        beam_size=args.beam_size,
+        schedule_type=args.schedule_type,
+        decay_a=args.decay_a,
+        decay_b=args.decay_b,
+        model_dir=args.model_output_dir)

scheduled_sampling/random_schedule_generator.py → scheduled_sampling/utils.py

-import numpy as np
 import math
+import numpy as np
 
 
 class RandomScheduleGenerator:
     """
-    The random sampling rate for scheduled sampling algoithm, which uses devcayed
+    The random sampling rate for scheduled sampling algoithm, which uses decayed
     sampling rate.
     """