基于循环神经网络的加法
欢迎指正。。
Keras实例目录
名词解释
循环神经网络(RNN)
长短期记忆网络(LSTM)
代码注释
# -*- coding: utf-8 -*-
'''An implementation of sequence to sequence learning for performing addition
基于相加的端到端学习实现
Input: "535+61"
输入:"535+61"
Output: "596"
输出:"596"
Padding is handled by using a repeated sentinel character (space)
使用重复的前哨字符(空格)填充
Input may optionally be inverted, shown to increase performance in many tasks in:
输入倒置(可选),在很多任务中展现提升的表现:
"Learning to Execute"
学习执行(论文)
http://arxiv.org/abs/1410.4615
and
"Sequence to Sequence Learning with Neural Networks"
基于神经网络的端到端学习
http://papers.nips.cc/paper/5346-sequence-to-sequence-learning-with-neural-networks.pdf
Theoretically it introduces shorter term dependencies between source and target.
理论上,引入了起源和目标之间的短期依赖关系。
Two digits inverted:
+ One layer LSTM (128 HN), 5k training examples = 99% train/test accuracy in 55 epochs
一层 LSTM(128 HN0,5k 训练样本 = 99% 训练/测试 精确度 50周期
Three digits inverted:
+ One layer LSTM (128 HN), 50k training examples = 99% train/test accuracy in 100 epochs
一层 LSTM(128 HN0,50k 训练样本 = 99% 训练/测试 精确度 100周期
Four digits inverted:
+ One layer LSTM (128 HN), 400k training examples = 99% train/test accuracy in 20 epochs
一层 LSTM(128 HN0,400k 训练样本 = 99% 训练/测试 精确度 20周期
Five digits inverted:
+ One layer LSTM (128 HN), 550k training examples = 99% train/test accuracy in 30 epochs
一层 LSTM(128 HN0,550k 训练样本 = 99% 训练/测试 精确度 30周期
'''
from __future__ import print_function
from keras.models import Sequential
from keras import layers
import numpy as np
from six.moves import range
class CharacterTable(object):
"""Given a set of characters:
+ Encode them to a one hot integer representation
one-hot 整数编码
+ Decode the one hot integer representation to their character output
one-hot 整数解码为字符输出
+ Decode a vector of probabilities to their character output
解码概率向量为字符输出
"""
def __init__(self, chars):
"""Initialize character table.
初始化字符表
# Arguments
参数
chars: Characters that can appear in the input.
字符:输入的字符
"""
self.chars = sorted(set(chars))
self.char_indices = dict((c, i) for i, c in enumerate(self.chars))
self.indices_char = dict((i, c) for i, c in enumerate(self.chars))
def encode(self, C, num_rows):
"""One hot encode given string C.
One hot编码处理字符串C
# Arguments
参数
num_rows: Number of rows in the returned one hot encoding. This is
used to keep the # of rows for each data the same.
num_rows: 用于one hot编码的行数,用来保持每行#数据相同
"""
x = np.zeros((num_rows, len(self.chars)))
for i, c in enumerate(C):
x[i, self.char_indices[c]] = 1
return x
def decode(self, x, calc_argmax=True):
if calc_argmax:
x = x.argmax(axis=-1)
return ''.join(self.indices_char[x] for x in x)
class colors:
ok = '\033[92m'
fail = '\033[91m'
close = '\033[0m'
# Parameters for the model and dataset.
# 模型和数据集参数
TRAINING_SIZE = 50000
DIGITS = 3
INVERT = True
# Maximum length of input is 'int + int' (e.g., '345+678'). Maximum length of
# int is DIGITS.
# 输入的最大长度是 'int + int' (例如 '345+678' (是3+1+3=7))。int 的最大长度是 DIGITS的值(本例是3)
MAXLEN = DIGITS + 1 + DIGITS # 7
# All the numbers, plus sign and space for padding.
# 所有数值、加号和填充的空格
chars = '0123456789+ ' # char: '0123456789+ '
ctable = CharacterTable(chars)
'''
CharacterTable基于chars = '0123456789+ '生成2个dict
char_indices = {dict} {' ': 0, '+': 1, '0': 2, '1': 3, '2': 4, '3': 5, '4': 6, '5': 7, '6': 8, '7': 9, '8': 10, '9': 11}
indices_char = {dict} {0: ' ', 1: '+', 2: '0', 3: '1', 4: '2', 5: '3', 6: '4', 7: '5', 8: '6', 9: '7', 10: '8', 11: '9'}
'''
questions = []
expected = []
seen = set()
print('Generating data...')
while len(questions) < TRAINING_SIZE:
f = lambda: int(''.join(np.random.choice(list('0123456789'))
for i in range(np.random.randint(1, DIGITS + 1))))
a, b = f(), f()
# Skip any addition questions we've already seen
# 跳过任何我们已经看过的附加问题
# Also skip any such that x+Y == Y+x (hence the sorting).
key = tuple(sorted((a, b)))
if key in seen:
continue
seen.add(key)
# Pad the data with spaces such that it is always MAXLEN.
# 用空格填充数据,使数据长度为最大长度(MAXLEN = DIGITS + 1 + DIGITS,是7)
q = '{}+{}'.format(a, b)
query = q + ' ' * (MAXLEN - len(q))
ans = str(a + b)
# Answers can be of maximum size DIGITS + 1.
# 答案可以是最大的 DIGITS + 1+ 1
ans += ' ' * (DIGITS + 1 - len(ans))
if INVERT:
# Reverse the query, e.g., '12+345 ' becomes ' 543+21'. (Note the
# space used for padding.)
# 翻转问题字符串(字符串字符位置倒置),例如 '12+345 ' 处理后 ' 543+21'(注意填充的空格(不要遗漏)。)
query = query[::-1]
questions.append(query)
expected.append(ans)
print('Total addition questions:', len(questions))
print('Vectorization...')
x = np.zeros((len(questions), MAXLEN, len(chars)), dtype=np.bool)
y = np.zeros((len(questions), DIGITS + 1, len(chars)), dtype=np.bool)
for i, sentence in enumerate(questions):
x[i] = ctable.encode(sentence, MAXLEN)
for i, sentence in enumerate(expected):
y[i] = ctable.encode(sentence, DIGITS + 1)
# Shuffle (x, y) in unison as the later parts of x will almost all be larger
# digits.
# 由于x数据集后面部分几乎都是较大的数字,同时筛选(x,y),即打乱数据集已有排序,但x、y对应关系不打乱。
indices = np.arange(len(y))
np.random.shuffle(indices)
x = x[indices]
y = y[indices]
# Explicitly set apart 10% for validation data that we never train over.
# 划分10%的数据(样本)作为验证数据集
split_at = len(x) - len(x) // 10
(x_train, x_val) = x[:split_at], x[split_at:]
(y_train, y_val) = y[:split_at], y[split_at:]
print('Training Data:')
print(x_train.shape)
print(y_train.shape)
print('Validation Data:')
print(x_val.shape)
print(y_val.shape)
# Try replacing GRU, or SimpleRNN.
# 尝试使用GRU(门控循环单元)或SimpleRNN替换
RNN = layers.LSTM
HIDDEN_SIZE = 128
BATCH_SIZE = 128
LAYERS = 1
print('Build model...')
model = Sequential()
# "Encode" the input sequence using an RNN, producing an output of HIDDEN_SIZE.
# Note: In a situation where your input sequences have a variable length,
# use input_shape=(None, num_feature).
# 使用RNN编码输入序列,生成HIDDEN_SIZE的输出
# 注意:某些情况,输入序列是变长的,使用input_shape=(None, num_feature).
model.add(RNN(HIDDEN_SIZE, input_shape=(MAXLEN, len(chars))))
# As the decoder RNN's input, repeatedly provide with the last hidden state of
# RNN for each time step. Repeat 'DIGITS + 1' times as that's the maximum
# length of output, e.g., when DIGITS=3, max output is 999+999=1998.
# 作为解码器RNN的输入,重复提供与最后一个隐藏状态的RNN为每个时间步长。重复“IGITS + 1”次,
# 这是输出的最大长度,例如,当DIGITS=3时,最大输出为999±999=1998。
model.add(layers.RepeatVector(DIGITS + 1))
# The decoder RNN could be multiple layers stacked or a single layer.
# 作为解码器的RNN可以是多层的或单层的。
for _ in range(LAYERS):
# By setting return_sequences to True, return not only the last output but
# all the outputs so far in the form of (num_samples, timesteps,
# output_dim). This is necessary as TimeDistributed in the below expects
# the first dimension to be the timesteps.
# 通过将返回序列设置为true,不仅返回最后的输出,而且还返回迄今为止的所有输出形式
# (NUMYSAMPLE,TimePosits,OutPuxDIM)。后续的时间分布需要以前的维度是时间步骤的。
model.add(RNN(HIDDEN_SIZE, return_sequences=True))
# Apply a dense layer to the every temporal slice of an input. For each of step
# of the output sequence, decide which character should be chosen.
# 输入的每一个时间切片上应用一个dense层。对于输出序列的每个步骤,决定应选择的字符。
model.add(layers.TimeDistributed(layers.Dense(len(chars))))
model.add(layers.Activation('softmax'))
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
model.summary()
# Train the model each generation and show predictions against the validation
# dataset.
# 训练模型并显示验证数据集的预测
for iteration in range(1, 200):
print()
print('-' * 50)
print('Iteration', iteration)
model.fit(x_train, y_train,
batch_size=BATCH_SIZE,
epochs=1,
validation_data=(x_val, y_val))
# Select 10 samples from the validation set at random so we can visualize
# errors.
# 从验证集随机选择10个样本,方便可视化错误
for i in range(10):
ind = np.random.randint(0, len(x_val))
rowx, rowy = x_val[np.array([ind])], y_val[np.array([ind])]
preds = model.predict_classes(rowx, verbose=0)
q = ctable.decode(rowx[0])
correct = ctable.decode(rowy[0])
guess = ctable.decode(preds[0], calc_argmax=False)
print('Q', q[::-1] if INVERT else q, end=' ')
print('T', correct, end=' ')
if correct == guess:
print(colors.ok + '☑' + colors.close, end=' ')
else:
print(colors.fail + '☒' + colors.close, end=' ')
print(guess)
class colors: ok = '\033[92m' fail = '\033[91m' close = '\033[0m'
该类用于标记验证效果,ok 为绿色;fail为红色,代码部分
print('Q', q[::-1] if INVERT else q, end=' ')
print('T', correct, end=' ')
if correct == guess:
print(colors.ok + '☑' + colors.close, end=' ')
else:
print(colors.fail + '☒' + colors.close, end=' ')
print(guess)
该类用于标记验证效果,ok 为绿色;fail为红色,效果如下
CharacterTable功能举例说明
执行过程
C:\ProgramData\Anaconda3\python.exe E:/keras-master/examples/addition_rnn.py
Using TensorFlow backend.
Generating data...
Total addition questions: 50000
Vectorization...
Training Data:
(45000, 7, 12)
(45000, 4, 12)
Validation Data:
(5000, 7, 12)
(5000, 4, 12)
Build model...
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
lstm_1 (LSTM) (None, 128) 72192
_________________________________________________________________
repeat_vector_1 (RepeatVecto (None, 4, 128) 0
_________________________________________________________________
lstm_2 (LSTM) (None, 4, 128) 131584
_________________________________________________________________
time_distributed_1 (TimeDist (None, 4, 12) 1548
_________________________________________________________________
activation_1 (Activation) (None, 4, 12) 0
=================================================================
Total params: 205,324
Trainable params: 205,324
Non-trainable params: 0
_________________________________________________________________
--------------------------------------------------
Iteration 1
Train on 45000 samples, validate on 5000 samples
Epoch 1/1
2018-02-19 22:44:31.792747: I C:\tf_jenkins\workspace\rel-win\M\windows\PY\36\tensorflow\core\platform\cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX
128/45000 [..............................] - ETA: 10:05 - loss: 2.4843 - acc: 0.0723
384/45000 [..............................] - ETA: 3:30 - loss: 2.4783 - acc: 0.1621
640/45000 [..............................] - ETA: 2:11 - loss: 2.4713 - acc: 0.1859
896/45000 [..............................] - ETA: 1:37 - loss: 2.4619 - acc: 0.1978
1152/45000 [..............................] - ETA: 1:18 - loss: 2.4510 - acc: 0.2036
1408/45000 [..............................] - ETA: 1:06 - loss: 2.4370 - acc: 0.2058
1664/45000 [>.............................] - ETA: 57s - loss: 2.4177 - acc: 0.2076
1920/45000 [>.............................] - ETA: 51s - loss: 2.3948 - acc: 0.2089
2176/45000 [>.............................] - ETA: 46s - loss: 2.3777 - acc: 0.2104
2432/45000 [>.............................] - ETA: 43s - loss: 2.3627 - acc: 0.2116
2688/45000 [>.............................] - ETA: 40s - loss: 2.3488 - acc: 0.2118
2944/45000 [>.............................] - ETA: 37s - loss: 2.3335 - acc: 0.2131
3200/45000 [=>............................] - ETA: 35s - loss: 2.3226 - acc: 0.2130
3456/45000 [=>............................] - ETA: 33s - loss: 2.3106 - acc: 0.2143
3712/45000 [=>............................] - ETA: 32s - loss: 2.3005 - acc: 0.2143
42880/45000 [===========================>..] - ETA: 0s - loss: 1.3089e-04 - acc: 1.0000
43136/45000 [===========================>..] - ETA: 0s - loss: 1.3101e-04 - acc: 1.0000
43392/45000 [===========================>..] - ETA: 0s - loss: 1.3093e-04 - acc: 1.0000
43648/45000 [============================>.] - ETA: 0s - loss: 1.3082e-04 - acc: 1.0000
43904/45000 [============================>.] - ETA: 0s - loss: 1.3075e-04 - acc: 1.0000
44160/45000 [============================>.] - ETA: 0s - loss: 1.3079e-04 - acc: 1.0000
44416/45000 [============================>.] - ETA: 0s - loss: 1.3073e-04 - acc: 1.0000
44672/45000 [============================>.] - ETA: 0s - loss: 1.3048e-04 - acc: 1.0000
44928/45000 [============================>.] - ETA: 0s - loss: 1.3037e-04 - acc: 1.0000
45000/45000 [==============================] - 10s 221us/step - loss: 1.3031e-04 - acc: 1.0000 - val_loss: 9.5955e-04 - val_acc: 0.9998
Q 478+869 T 1347 ☑ 1347
Q 36+34 T 70 ☑ 70
Q 303+382 T 685 ☑ 685
Q 1+611 T 612 ☑ 612
Q 674+72 T 746 ☑ 746
Q 803+32 T 835 ☑ 835
Q 444+948 T 1392 ☑ 1392
Q 634+34 T 668 ☑ 668
Q 4+208 T 212 ☑ 212
Q 8+970 T 978 ☑ 978
Process finished with exit code 0