fix tiny errors

code/models.py

@@ -3,9 +3,6 @@
 import numpy as np
 
 class SiameseNN(object):
-    """
-    pairwise学习模型
-    """
     def __init__(self, config):
         self.config = config
         # 输入
@@ -102,9 +99,6 @@ def add_train_op(self, loss):
 
 
 class SiameseCNN(object):
-    """
-    pairwise学习模型
-    """
     def __init__(self, config):
         self.config = config
         # 输入
@@ -113,7 +107,7 @@ def __init__(self, config):
         q_embed, a_embed = self.add_embeddings()
         with tf.variable_scope('siamese') as scope:
             self.q_trans = self.network(q_embed, reuse=False)
-            tf.get_variable_scope().reuse_variables()
+            scope.reuse_variables()
             self.a_trans = self.network(a_embed, reuse=True)
         # 损失和精确度
         self.total_loss = self.add_loss_op(self.q_trans, self.a_trans)
@@ -163,7 +157,7 @@ def fc_layer(self, bottom, n_weight, name):
         n_prev_weight = bottom.get_shape()[1]
         initer = tf.truncated_normal_initializer(stddev=0.01)
         W = tf.get_variable(name+'W', dtype=tf.float32, shape=[n_prev_weight, n_weight], initializer=initer)
-        b = tf.get_variable(name+'b', dtype=tf.float32, initializer=tf.constant(0.01, shape=[n_weight], dtype=tf.float32))
+        b = tf.get_variable(name+'b', dtype=tf.float32, initializer=tf.constant(0.0, shape=[n_weight], dtype=tf.float32))
         fc = tf.nn.bias_add(tf.matmul(bottom, W), b)
         return fc
 
@@ -173,10 +167,10 @@ def conv_layer(self, h, reuse=False):
         h = tf.reshape(h, [-1, max_len, h.get_shape()[2], 1])
         for i, filter_size in enumerate(self.config.filter_sizes):
             with tf.variable_scope('filter{}'.format(filter_size)):
-                conv1_W = tf.get_variable('conv_W', shape=[filter_size, self.config.embedding_size, 1, self.config.num_filters], initializer=tf.truncated_normal_initializer(.0, .1))
-                conv1_b = tf.get_variable('conv_b', initializer=tf.constant(0.1, shape=[self.config.num_filters]))
+                conv1_W = tf.get_variable('conv_W', shape=[filter_size, self.config.embedding_size, 1, self.config.num_filters], initializer=tf.truncated_normal_initializer(.0, .01))
+                conv1_b = tf.get_variable('conv_b', initializer=tf.constant(0.0, shape=[self.config.num_filters]))
                 # pooling层的bias,Q和A分开
-                pool_b = tf.get_variable('pool_b', initializer=tf.constant(0.1, shape=[self.config.num_filters]))
+                pool_b = tf.get_variable('pool_b', initializer=tf.constant(0.0, shape=[self.config.num_filters]))
                 # 卷积
                 out = tf.nn.relu((tf.nn.conv2d(h, conv1_W, [1,1,1,1], padding='VALID')+conv1_b))
                 # 池化
@@ -185,10 +179,10 @@ def conv_layer(self, h, reuse=False):
                 pool.append(out)
                 # 加入正则项
                 if not reuse:
-                    tf.add_to_collection('total_loss', 0.5*self.config.l2_reg_lambda*tf.nn.l2_loss(conv1_W))
+                    tf.add_to_collection('total_loss', 0.5 * self.config.l2_reg_lambda * tf.nn.l2_loss(conv1_W))
 
-        total_channels = len(self.config.filter_sizes)*self.config.num_filters
-        real_pool = tf.reshape(tf.concat(pool, 3), [-1, total_channels])
+        total_channels = len(self.config.filter_sizes) * self.config.num_filters
+        real_pool = tf.reshape(tf.concat(pool, 3), [self.batch_size, total_channels])
         return real_pool
 
     # 损失节点
@@ -220,9 +214,6 @@ def add_train_op(self, loss):
 
 
 class SiameseRNN(object):
-    """
-    pairwise学习模型
-    """
     def __init__(self, config):
         self.config = config
         # 输入
@@ -272,7 +263,7 @@ def network(self, x):
         inputs = tf.transpose(x, [1, 0, 2])
         inputs = tf.reshape(inputs, [-1, self.config.embedding_size])
         inputs = tf.split(inputs, sequence_length, 0)
-        # (batch_size, conv_size)
+        # (batch_size, rnn_output_size)
         rnn1 = self.rnn_layer(inputs)
         # (batch_size, hidden_size)
         fc1 = self.fc_layer(rnn1, self.config.hidden_size, "fc1")
@@ -494,4 +485,4 @@ def add_train_op(self, loss):
             self.global_step = tf.Variable(0, name='global_step', trainable=False)
             opt = tf.train.AdamOptimizer(self.config.lr)
             train_op = opt.minimize(loss, self.global_step)
-            return train_op
+            return train_op

code/preprocess_wiki.ipynb

@@ -74,6 +74,7 @@
     }
    ],
    "source": [
+    "# 分词、词干化处理\n",
     "def segment(filename, use_lemma=True):\n",
     "    processed_qa = []\n",
     "    count = 0\n",
@@ -94,6 +95,7 @@
     "                print('Finished {}'.format(count))\n",
     "    return processed_qa\n",
     "\n",
+    "# 构建词典\n",
     "def build_vocab(corpus, topk=None):\n",
     "    vocab = Counter()\n",
     "    for line in corpus:\n",
@@ -110,6 +112,7 @@
     "    reverse_vocab = dict(zip(vocab.values(), vocab.keys()))\n",
     "    return vocab, reverse_vocab\n",
     "\n",
+    "# 将每个词映射为词典中的id\n",
     "def transform(corpus, word2id, unk_id=1):\n",
     "    transformed_corpus = []\n",
     "    for line in corpus:\n",
@@ -119,6 +122,7 @@
     "        transformed_corpus.append([qid, q, aid, a, int(label)])\n",
     "    return transformed_corpus\n",
     "\n",
+    "# 得到pointwise形式的数据，即(Q, A, label)\n",
     "def pointwise_data(corpus, keep_ids=False):\n",
     "    # (q, a, label)\n",
     "    pointwise_corpus = []\n",
@@ -130,6 +134,7 @@
     "            pointwise_corpus.append((q, a, label))\n",
     "    return pointwise_corpus\n",
     "\n",
+    "# 得到pairwise形式的数据，即(Q, positive A, negative A)\n",
     "def pairwise_data(corpus):\n",
     "    # (q, a_pos, a_neg), two answers must from the same q\n",
     "    # once a question contains no positive answers, we discard this sample.\n",
@@ -152,6 +157,7 @@
     "                real_pairwise_corpus.append((q, pos, neg))\n",
     "    return real_pairwise_corpus\n",
     "    \n",
+    "# 得到listwise形式的数据，即(Q, All answers related to this Q)\n",
     "def listwise_data(corpus):\n",
     "    # (q, a_list)\n",
     "    listwise_corpus = dict()\n",
@@ -168,7 +174,7 @@
     "        real_listwise_corpus.append((q, alist))\n",
     "    return real_listwise_corpus\n",
     "\n",
-    "            \n",
+    "\n",
     "train_processed_qa = segment('WikiQA-train.tsv')\n",
     "val_processed_qa = segment('WikiQA-dev.tsv')\n",
     "test_processed_qa = segment('WikiQA-test.tsv')\n",

code/siamese.py

@@ -3,6 +3,7 @@
 import numpy as np
 import os
 import sys
+from copy import deepcopy
 stdout = sys.stdout
 reload(sys)
 sys.stdout = stdout
@@ -61,15 +62,15 @@ def __init__(self, vocab_size, embeddings=None):
         self.embedding_size = 100
         if self.embeddings is not None:
             self.embedding_size = embeddings.shape[1]
-        # 不同类型的filter,相当于1-gram,2-gram,3-gram和5-gram
+        # 不同类型的filter，对应不同的尺寸
         self.filter_sizes = [1, 2, 3, 5, 7, 9]
         # 隐层大小
         self.hidden_size = 256
         self.output_size = 128
         # 每种filter的数量
         self.num_filters = 128
         self.l2_reg_lambda = 0.
-        self.keep_prob = 0.6
+        self.keep_prob = 0.8
         # 学习率
         self.lr = 0.001
         # contrasive loss 中的 positive loss部分的权重
@@ -105,7 +106,6 @@ def __init__(self, vocab_size, embeddings=None):
         self.output_size = 128
         # 每种filter的数量
         self.num_filters = 128
-        self.l2_reg_lambda = 0.
         self.keep_prob = 0.6
         # 学习率
         self.lr = 0.001
@@ -142,7 +142,7 @@ def train(train_corpus, config, val_corpus, eval_train_corpus=None):
                 count += 1
                 if count % 10 == 0:
                     print('[epoch {}, batch {}]Loss:{}'.format(epoch, count, loss))
-            saver.save(sess,'models/siamese_nn/my_model', global_step=epoch)
+            saver.save(sess,'models/siamese_{}/my_model'.format(args.model.lower()), global_step=epoch)
             if eval_train_corpus is not None:
                 train_res = evaluate(sess, model, eval_train_corpus, config)
                 print('[train] ' + train_res)
@@ -195,7 +195,7 @@ def test(corpus, config):
         else:
             model = SiameseRNN(config)
         saver = tf.train.Saver()
-        saver.restore(sess, tf.train.latest_checkpoint('models/qacnn'))
+        saver.restore(sess, tf.train.latest_checkpoint('models/siamese_{}'.format(args.model)))
         print('[test] ' + evaluate(sess, model, corpus, config))
 
 
@@ -234,7 +234,7 @@ def main(args):
     config.max_q_length = max_q_length
     config.max_a_length = max_a_length
     if args.train:
-        train(train_corpus, config, val_corpus, train_corpus)
+        train(deepcopy(train_corpus), config, val_corpus, deepcopy(train_corpus))
     elif args.test:
         test(test_corpus, config)