reference:https://blog.csdn.net/aliceyangxi1987/article/details/73420844
我自己在ubuntu上运行完，内存报了个错，应该是机器的问题，忽略： Epoch 598 Epoch 599 Traceback (most recent call last): File "train.py", line 77, in incorrect = sess.run(error,{data: test_input, target: test_output}) File "/home/wqu/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 900, in run run_metadata_ptr) File "/home/wqu/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 1104, in _run np_val = np.asarray(subfeed_val, dtype=subfeed_dtype) File "/home/wqu/.local/lib/python2.7/site-packages/numpy/core/numeric.py", line 492, in asarray return array(a, dtype, copy=False, order=order) MemoryError
用一个简单的例子来看看 LSTM 在 tensorflow 里是如何做分类问题的。
这个例子特别简单，就是一个长度为 20 的二进制串，数出其中 1 的个数，简单到用一个 for 就能搞定的事情，来看看 LSTM 是如何做到的。
大家可以先在这里停一下，看看你有什么想法呢。
import tensorflow as tf import numpy as np from random import shuffle12
input 一共有 2^20 种组合，就生成这么多的数据
train_input = ['{0:020b}'.format(i) for i in range(2**20)] shuffle(train_input) train_input = [map(int,i) for i in train_input]123
train_input： [1, 0, 0, 0, 1, 1, 1, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0] [0, 0, 0, 1, 1, 0, 1, 1, 1, 0, 1, 1, 0, 1, 0, 1, 0, 1, 0, 1] [0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 1, 0, 0, 1, 1, 0, 0, 1]
把每一个 input 转化成 tensor 的形式 在 dimensions ＝ [batch_size, sequence_length, input_dimension] 中， sequence_length ＝ 20 and input_dimension ＝ 1， 每个 input 变成了 A list of 20 lists
ti = []
for i in train_input:
temp_list = []
for j in i:
temp_list.append([j])
ti.append( np.array(temp_list) )
train_input = ti12345678
train_input ： [[1][0][0][0][1][1][1][0][1][0][0][0][0][1][0][0][0][1][0][0]]
生成实际的 output 数据
train_output = []
for i in train_input: count = 0 for j in i: if j[0] == 1: count+=1 temp_list = ([0]*21) temp_list[count]=1 train_output.append(temp_list)12345678910
train_output：在第几个位置上有一个 1 ，说明 input 里面就有几个 1，长度为 21 [0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0] [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0]
取 0.9% 为训练数据，另外的为测试数据
NUM_EXAMPLES = 10000 test_input = train_input[NUM_EXAMPLES:] test_output = train_output[NUM_EXAMPLES:] #everything beyond 10,000
train_input = train_input[:NUM_EXAMPLES] train_output = train_output[:NUM_EXAMPLES] #till 10,000123456
定义两个变量 其中 data 的维度 ＝ [Batch Size, Sequence Length, Input Dimension]
data = tf.placeholder(tf.float32, [None, 20,1]) target = tf.placeholder(tf.float32, [None, 21])12
定义 hidden dimension ＝ 24 太多会 overfitting，太少效果不好，可以调节看变化。 模型用 LSTM，这里用的 tf 1.0.0 的 version
num_hidden = 24
cell = tf.nn.rnn_cell.LSTMCell(num_hidden,state_is_tuple=True)
cell = tf.contrib.rnn.LSTMCell(num_hidden,state_is_tuple=True)123
用 val 来存这个 output
val, _ = tf.nn.dynamic_rnn(cell, data, dtype=tf.float32)1
变换一下维度，并取 val 的最后一个为 last
val = tf.transpose(val, [1, 0, 2]) last = tf.gather(val, int(val.get_shape()[0]) - 1)12
定义 weight 和 bias
weight = tf.Variable(tf.truncated_normal( [num_hidden, int(target.get_shape()[1])] )) bias = tf.Variable(tf.constant(0.1, shape=[target.get_shape()[1]]))12
再作用上 softmax 得到 prediction
prediction = tf.nn.softmax(tf.matmul(last, weight) + bias)1
用 cross_entropy 来做 cost function，目标是使它最小化，选用 AdamOptimizer
cross_entropy = -tf.reduce_sum(target * tf.log(tf.clip_by_value(prediction,1e-10,1.0)))
optimizer = tf.train.AdamOptimizer() minimize = optimizer.minimize(cross_entropy)1234
定义一下 error 的形式，就是预测和实际有多少个位置不一样
mistakes = tf.not_equal(tf.argmax(target, 1), tf.argmax(prediction, 1)) error = tf.reduce_mean(tf.cast(mistakes, tf.float32))12
前面定义完模型和变量，这里开始启动 session
init_op = tf.initialize_all_variables() sess = tf.Session() sess.run(init_op)123
迭代 600 次就可以达到 0.3% 的 error 了
batch_size = 1000 no_of_batches = int(len(train_input)) / batch_size epoch = 600123
for i in range(epoch): ptr = 0 for j in range(no_of_batches): inp, out = train_input[ptr:ptr+batch_size], train_output[ptr:ptr+batch_size] ptr += batch_size sess.run(minimize,{data: inp, target: out}) print "Epoch ",str(i)
incorrect = sess.run(error,{data: test_input, target: test_output})
print sess.run(prediction, {data: [[[1],[0],[0],[1],[1],[0],[1],[1],[1],[0],[1],[0],[0],[1],[1],[0],[1],[1],[1],[0]]]}) print('Epoch {:2d} error {:3.1f}%'.format(i + 1, 100 * incorrect))
sess.close()1234567891011121314
最后的结果：
[[ 2.80220238e-08 3.24575727e-10 5.68697936e-11 3.57573054e-10 9.62089857e-08 1.30921896e-08 2.14473985e-08 5.21751364e-10 2.29034747e-08 8.47907577e-10 3.60394756e-06 2.30961153e-03 9.82593179e-01 1.50928665e-02 4.23395448e-07 1.06428047e-07 6.70640388e-09 1.78888765e-10 3.22445395e-08 3.09186134e-08 3.70296416e-09]]
Epoch 600 error 0.3%