Merge remote-tracking branch 'upstream/develop' into restructure_ltr

conv_seq2seq/beamsearch.py

@@ -11,7 +11,6 @@ import reader
 class BeamSearch(object):
     """
     Generate sequence by beam search
-    NOTE: this class only implements generating one sentence at a time.
     """
 
     def __init__(self,

generate_sequence_by_rnn_lm/network_conf.py

@@ -57,4 +57,4 @@ def rnn_lm(vocab_dim,
     else:
         cost = paddle.layer.classification_cost(input=output, label=target)
 
-        return cost, output
+        return cost

generate_sequence_by_rnn_lm/train.py

@@ -43,9 +43,14 @@ def train(topology,
 
     # create parameters
     parameters = paddle.parameters.create(topology)
+    # create sum evaluator
+    sum_eval = paddle.evaluator.sum(topology)
     # create trainer
     trainer = paddle.trainer.SGD(
-        cost=topology, parameters=parameters, update_equation=adam_optimizer)
+        cost=topology,
+        parameters=parameters,
+        update_equation=adam_optimizer,
+        extra_layers=sum_eval)
 
     # define the event_handler callback
     def event_handler(event):

nested_sequence/text_classification/network_conf.py

@@ -15,11 +15,13 @@ def cnn_cov_group(group_input, hidden_size):
     conv4 = paddle.networks.sequence_conv_pool(
         input=group_input, context_len=4, hidden_size=hidden_size)
 
+    fc_param_attr = paddle.attr.ParamAttr(name='_cov_value_weight')
+    fc_bias_attr = paddle.attr.ParamAttr(name='_cov_value_bias')
     linear_proj = paddle.layer.fc(
         input=[conv3, conv4],
         size=hidden_size,
-        param_attr=paddle.attr.ParamAttr(name='_cov_value_weight'),
-        bias_attr=paddle.attr.ParamAttr(name='_cov_value_bias'),
+        param_attr=[fc_param_attr, fc_param_attr],
+        bias_attr=fc_bias_attr,
         act=paddle.activation.Linear())
 
     return linear_proj

sequence_tagging_for_ner/network_conf.py

@@ -8,7 +8,7 @@ def ner_net(word_dict_len, label_dict_len, stack_num=2, is_train=True):
     mark_dict_len = 2
     word_dim = 50
     mark_dim = 5
-    hidden_dim = 128
+    hidden_dim = 300
 
     word = paddle.layer.data(
         name="word",
@@ -23,9 +23,7 @@ def ner_net(word_dict_len, label_dict_len, stack_num=2, is_train=True):
         name="mark",
         type=paddle.data_type.integer_value_sequence(mark_dict_len))
     mark_embedding = paddle.layer.embedding(
-        input=mark,
-        size=mark_dim,
-        param_attr=paddle.attr.Param(initial_std=math.sqrt(1. / word_dim)))
+        input=mark, size=mark_dim, param_attr=paddle.attr.Param(initial_std=0.))
 
     word_caps_vector = paddle.layer.concat(
         input=[word_embedding, mark_embedding])
@@ -33,7 +31,7 @@ def ner_net(word_dict_len, label_dict_len, stack_num=2, is_train=True):
     mix_hidden_lr = 1e-3
     rnn_para_attr = paddle.attr.Param(initial_std=0.0, learning_rate=0.1)
     hidden_para_attr = paddle.attr.Param(
-        initial_std=1 / math.sqrt(hidden_dim), learning_rate=mix_hidden_lr)
+        initial_std=1. / math.sqrt(hidden_dim) / 3, learning_rate=mix_hidden_lr)
 
     # the first forward and backward rnn layer share the
     # input-to-hidden mappings.
@@ -41,9 +39,10 @@ def ner_net(word_dict_len, label_dict_len, stack_num=2, is_train=True):
         name="__hidden00__",
         size=hidden_dim,
         act=paddle.activation.Tanh(),
-        bias_attr=paddle.attr.Param(initial_std=1.),
+        bias_attr=paddle.attr.Param(initial_std=1. / math.sqrt(hidden_dim) / 3),
         input=word_caps_vector,
-        param_attr=hidden_para_attr)
+        param_attr=paddle.attr.Param(initial_std=1. / math.sqrt(hidden_dim) /
+                                     3))
 
     fea = []
     for direction in ["fwd", "bwd"]:
@@ -68,7 +67,7 @@ def ner_net(word_dict_len, label_dict_len, stack_num=2, is_train=True):
 
     rnn_fea = paddle.layer.fc(
         size=hidden_dim,
-        bias_attr=paddle.attr.Param(initial_std=1.),
+        bias_attr=paddle.attr.Param(initial_std=1. / math.sqrt(hidden_dim) / 3),
         act=paddle.activation.STanh(),
         input=fea,
         param_attr=[hidden_para_attr, rnn_para_attr] * 2)
@@ -85,7 +84,8 @@ def ner_net(word_dict_len, label_dict_len, stack_num=2, is_train=True):
         bias_attr=False,
         input=rnn_fea,
         act=paddle.activation.Linear(),
-        param_attr=rnn_para_attr)
+        param_attr=paddle.attr.Param(initial_std=1. / math.sqrt(hidden_dim) /
+                                     3))
 
     if is_train:
         target = paddle.layer.data(
@@ -96,7 +96,10 @@ def ner_net(word_dict_len, label_dict_len, stack_num=2, is_train=True):
             size=label_dict_len,
             input=emission,
             label=target,
-            param_attr=paddle.attr.Param(name="crfw", initial_std=1e-3))
+            param_attr=paddle.attr.Param(
+                name="crfw",
+                initial_std=1. / math.sqrt(hidden_dim) / 3,
+                learning_rate=mix_hidden_lr))
 
         crf_dec = paddle.layer.crf_decoding(
             size=label_dict_len,

sequence_tagging_for_ner/train.py

@@ -16,8 +16,8 @@ def main(train_data_file,
          target_file,
          emb_file,
          model_save_dir,
-         num_passes=10,
-         batch_size=32):
+         num_passes=100,
+         batch_size=64):
     if not os.path.exists(model_save_dir):
         os.mkdir(model_save_dir)
 
@@ -75,10 +75,10 @@ def main(train_data_file,
 
     def event_handler(event):
         if isinstance(event, paddle.event.EndIteration):
-            if event.batch_id % 1 == 0:
+            if event.batch_id % 5 == 0:
                 logger.info("Pass %d, Batch %d, Cost %f, %s" % (
                     event.pass_id, event.batch_id, event.cost, event.metrics))
-            if event.batch_id % 1 == 0:
+            if event.batch_id % 50 == 0:
                 result = trainer.test(reader=test_reader, feeding=feeding)
                 logger.info("\nTest with Pass %d, Batch %d, %s" %
                             (event.pass_id, event.batch_id, result.metrics))