fix test_atten.py; update readme;

README.md

@@ -127,6 +127,10 @@ Input English Sentence:i'm really a bad boy
 
 [Chatbot实例](/chatbot/)
 
+
+`test_atten.py` 脚本，测试并展示 attention 的热力图
+
+
 # TensorFlow alert
 
 Test in

chatbot_cut/read_vector.py

@@ -1,11 +1,32 @@
+"""
+读取一个文本格式的，保存预训练好的embedding的文件
 
+wiki.zh.vec
+
+它的第一行会被忽略
+第二行开始，每行是 词 + 空格 + 词向量维度0 + 空格 + 词向量维度1 + ...
+
+参考fasttext的文本格式
+
+https://github.com/facebookresearch/fastText/blob/master/pretrained-vectors.md
+"""
 
 import pickle
 import numpy as np
 from tqdm import tqdm
 
 
 def read_vector(path='wiki.zh.vec', output_path='word_vec.pkl'):
+    """
+    读取文本文件 path 中的数据，并且生成一个 dict 写入到 output_path
+
+    格式：
+    word_vec = {
+        'word_1': np.array(vec_of_word_1),
+        'word_2': np.array(vec_of_word_2),
+        ...
+    }
+    """
     fp = open(path, 'r')
     word_vec = {}
     first_skip = False

sequence_to_sequence.py

@@ -25,7 +25,7 @@
 import numpy as np
 import tensorflow as tf
 from tensorflow import layers
-from tensorflow.python.util import nest
+# from tensorflow.python.util import nest
 from tensorflow.python.ops import array_ops
 from tensorflow.contrib import seq2seq
 from tensorflow.contrib.seq2seq import BahdanauAttention
@@ -37,7 +37,7 @@
 from tensorflow.contrib.rnn import MultiRNNCell
 from tensorflow.contrib.rnn import DropoutWrapper
 from tensorflow.contrib.rnn import ResidualWrapper
-from tensorflow.contrib.rnn import LSTMStateTuple
+# from tensorflow.contrib.rnn import LSTMStateTuple
 
 from word_sequence import WordSequence
 from data_utils import _get_embed_device
@@ -414,8 +414,6 @@ def build_encoder(self):
                     parallel_iterations=self.parallel_iterations,
                     swap_memory=True
                 )
-
-                return encoder_outputs, encoder_state
             else:
                 # 双向 RNN 比较麻烦
                 encoder_cell_bw = self.build_encoder_cell()
@@ -443,7 +441,7 @@ def build_encoder(self):
                     encoder_state.append(encoder_bw_state[i])
                 encoder_state = tuple(encoder_state)
 
-                return encoder_outputs, encoder_state
+            return encoder_outputs, encoder_state
 
 
     def build_decoder_cell(self, encoder_outputs, encoder_state):
@@ -499,7 +497,7 @@ def build_decoder_cell(self, encoder_outputs, encoder_state):
 
         # 在非训练（预测）模式，并且没开启 beamsearch 的时候，打开 attention 历史信息
         alignment_history = (
-            not self.mode == 'train' and not self.use_beamsearch_decode
+            self.mode != 'train' and not self.use_beamsearch_decode
         )
 
         def cell_input_fn(inputs, attention):
@@ -1091,7 +1089,7 @@ def predict(self, sess,
 
             pred, atten = sess.run([
                 self.decoder_pred_decode,
-                self.final_state[1].alignment_history.stack()
+                self.final_state.alignment_history.stack()
             ], input_feed)
 
             return pred, atten

test_atten.py

@@ -22,7 +22,7 @@ def test(bidirectional, cell_type, depth, attention_type):
 
     # 训练部分
 
-    split = 9900
+    split = int(len(x_data) * 0.9)
     x_train, x_test, y_train, y_test = (
         x_data[:split], x_data[split:], y_data[:split], y_data[split:])
     n_epoch = 2
@@ -86,7 +86,6 @@ def test(bidirectional, cell_type, depth, attention_type):
             cell_type=cell_type,
             depth=depth,
             attention_type=attention_type,
-            alignment_history=True,
             parallel_iterations=1
         )
         init = tf.global_variables_initializer()