我想在Tensorflow中使用这个递归神经网络(来自本教程
)https://github.com/aymericdamien/TensorFlow-Examples/
然后是RNN程序))，这样它将是一个去噪的自动编码器。我有5个时间步，每次，无噪声的目标从sin(x)采样，有噪声的输入是sin(x)+高斯误差。现在我的问题是，示例中的RNN为每个输入序列提供了1个输出值，但我想要每个时间步的输出(我想要5个输出，而不是1个)我该怎么做呢?我怀疑这可能是重新定义权重和偏差的问题，但怎么做呢?下面是代码。非常感谢你的帮助，
import tensorflow as tf
from tensorflow.python.ops import rnn, rnn_cell
import numpy as np
# Parameters
learning_rate = 0.0005
training_iters = 1000
batch_size = 3
display_step = 100
# Network Parameters
n_input = 2
n_output = 2
n_steps = 5 # timesteps
n_hidden = 40 # hidden layer num of features
# tf Graph input
x = tf.placeholder("float", [None, n_steps, n_input])
y = tf.placeholder("float", [None, n_steps, n_input])
# Define weights
weights = {
'out': tf.Variable(tf.random_normal([n_hidden, n_output]))
}
biases = {
'out': tf.Variable(tf.random_normal([ n_output]))
}
# length of time series to be sampled
N = 1000000
dim_input = 2
x1 = np.zeros(N)
x2 = np.zeros(N)
y1 = np.zeros(N)
y2 = np.zeros(N)
# generate data
for i in range(0,N):
# clean
y1[i] = np.math.sin(i)
y2[i] = np.math.cos(i)
# noisy
x1[i] = y1[i]+np.random.normal(loc=0.0, scale=0.05)
x2[i] = y2[i]+np.random.normal(loc=0.0, scale=0.05)
def next_batch():
batch = np.empty([batch_size,n_steps,dim_input])
batch_y = np.empty([batch_size,n_steps,dim_input])
# for plotting purposes only
inits = np.empty([batch_size], dtype=int)
for b in range(0,batch_size):
# the first one of the batch
inits[b] = int(np.round(np.random.uniform(low=0,high=N-n_steps- 1)))
init = inits[b]
for i in range(0,n_steps):
# noisy input
batch[b,i,0] = x1[init + i]
batch[b,i,1] = x2[init + i]
# target (no noise)"
batch_y[b,i,0] = y1[init+i]
batch_y[b,i,1] = y2[init+i]
return(batch,batch_y,inits)
def RNN(x, weights, biases):
x = tf.transpose(x, [1, 0, 2])
# Reshaping to (n_steps*batch_size, n_input)
x = tf.reshape(x, [-1, n_input])
# Split to get a list of 'n_steps' tensors of shape (batch_size, n_input)
x = tf.split(0, n_steps, x)
# Define a lstm cell with tensorflow
lstm_cell = rnn_cell.BasicLSTMCell(n_hidden, forget_bias=1.0)
# Get lstm cell output
outputs, states = rnn.rnn(lstm_cell, x, dtype=tf.float32)
# Linear activation, using rnn inner loop last output
return tf.matmul(outputs[-1], weights['out']) + biases['out']
pred = RNN(x, weights, biases)
# Define loss and optimizer
# SSE, there must be an easier way to do this
def get_cost(prediction,truth):
z = 0
for i in range(0,batch_size):
z = z + np.square(np.add(prediction[i,:], np.multiply(-1,truth[i,:])))
z = np.add(z[0],z[1])
z = np.sum(z)
return(z)
cost = get_cost(pred,y)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).
minimize(cost)
# Evaluate model
accuracy = cost
# Initializing the variables
init = tf.initialize_all_variables()
# Launch the graph
with tf.Session() as sess:
sess.run(init)
step = 1
# Keep training until reach max iterations
while step * batch_size < training_iters:
print('step '+ str(step))
batch_x, batch_y, inits = next_batch()
sess.run(optimizer, feed_dict={x: batch_x, y: batch_y})
if step % display_step == 0:
# Calculate batch accuracy
acc = sess.run(accuracy, feed_dict={x: batch_x, y: batch_y})
# Calculate batch loss
loss = sess.run(cost, feed_dict={x: batch_x, y: batch_y})
print(loss)
step += 1
print("Optimization Finished!")
如果我运行这个，我得到这个错误消息:错误:Shape (?)， 5,2)必须是第2位。这看起来很合理，因为目标有5步长，而输出只有1步。但我该如何解决这个问题呢?
许多谢谢。import tensorflow as tf
from tensorflow.python.ops import rnn, rnn_cell
import numpy as np
import matplotlib.pyplot as plt
## Denoising autoencoder.
import numpy as np
count = 0
# length of time series to be sampled
N = 10000
x1 = np.zeros(N)
x2 = np.zeros(N)
y1 = np.zeros(N)
y2 = np.zeros(N)
batch_size = 30
learning_rate = 0.0005
training_iters = 300000
display_step = 100
# Network Parameters
n_input = 2
n_output = 2
n_steps = 15 # timesteps
n_hidden = 75 # hidden layer num of
# generate data
for i in range(0,N):
# clean
y1[i] = np.math.sin(i)
y2[i] = np.math.cos(i)
# noisy
x1[i] = y1[i]+np.random.normal(loc=0.0, scale=0.1)
x2[i] = y2[i]+np.random.normal(loc=0.0, scale=0.1)
def next_batch():
batch = np.empty([batch_size,n_steps,n_input])
batch_y = np.empty([batch_size,n_steps,n_input])
# for plotting purposes only
inits = np.empty([batch_size], dtype=int)
for b in range(0,batch_size):
# the first one of the batch
inits[b] = int(np.round(np.random.uniform(low=0,high=N-n_steps-1)))
init = inits[b]
for i in range(0,n_steps):
# noisy input
batch[b,i,0] = x1[init + i]
batch[b,i,1] = x2[init + i]
# target (no noise)"
batch_y[b,i,0] = y1[init+i]
batch_y[b,i,1] = y2[init+i]
return(batch,batch_y,inits)
# Parameters
# tf Graph input
x = tf.placeholder("float", [None, n_steps, n_input])
y = tf.placeholder("float", [None, n_steps, n_output])
N_train = N - 500
def RNN(x):
# Prepare data shape to match `rnn` function requirements
# Current data input shape: (batch_size, n_steps, n_input)
# Required shape: 'n_steps' tensors list of shape (batch_size, n_input)
# Permuting batch_size and n_steps
x = tf.transpose(x, [1, 0, 2])
# Reshaping to (n_steps*batch_size, n_input)
x = tf.reshape(x, [-1, n_input])
# Split to get a list of 'n_steps' tensors of shape (batch_size, n_input)
x = tf.split(0, n_steps, x)
# Define a lstm cell with tensorflow
lstm_cell = rnn_cell.LSTMCell(num_units = n_hidden, forget_bias=1.0, num_proj=2)
# Get lstm cell output
outputs, states = rnn.rnn(lstm_cell, x, dtype=tf.float32)
return outputs
print(x)
pred = RNN(x)
# Define loss and optimizer
def get_cost(prediction,truth):
#print('pred' + str(prediction))
# SSE. there must be an easier way than this:
z = 0
for step in range(0,n_steps):
for b in range(0,batch_size):
for y_dim in range(0,2):
d1 = prediction[step][b,y_dim]
d2 = truth[b,step,y_dim]
diff= (d1 - d2 )
z = z + diff * diff
return(z)
cost = get_cost(pred,y)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)
# Evaluate model
# Initializing the variables
init = tf.initialize_all_variables()
# Launch the graph
with tf.Session() as sess:
sess.run(init)
step = 1
# Keep training until reach max iterations
while step * batch_size < training_iters:
#print('step '+ str(step))
batch_x, batch_y, inits = next_batch()
# Reshape data to get 28 seq of 28 elements
#batch_x = batch_x.reshape((batch_size, n_steps, n_input))
# Run optimization op (backprop)
sess.run(optimizer, feed_dict={x: batch_x, y: batch_y})
if step % display_step == 0:
# Calculate batch loss
loss = sess.run(cost, feed_dict={x: batch_x, y: batch_y})
print(str(step) + ':' + str(loss))
step += 1
print("Optimization Finished!")
batch_size = 1
test_data, test_label, inits = next_batch()
#print "Testing Accuracy:",
#sess.run(accuracy, feed_dict={x: test_data, y: test_label})
p2 = sess.run(pred, feed_dict={x: test_data, y: test_label})
#print('---inits---')
#print(inits)
print('---batch---')
print(test_data)
print('---truth---')
print(test_label)
print('---pred---')
print(p2)
c_final = get_cost(p2, test_label)
print(c_final)
首先，我们生成一些数据:sin(i)和cos(i)的二维序列，其中i从1到n运行，这给了我们变量y。然后我们向该序列添加一些正常噪声，这是x。然后，我们训练一个循环神经网络来从噪声输入中创建干净的输出。换句话说，我们训练网络，使其从输入[cos(i)+e1,sin(i)+e2]输出[cos(i)，sin(i)]。这是一个普通的去噪自动编码器，除了数据有一个时间元素。现在你可以把新数据输入神经网络，它有望消除噪音。