tf.reshape没有按预期工作

时间：2017-02-07 06:11:46

标签： tensorflow

实际上，我不知道如何描述这个问题。太奇怪了。

import tensorflow as tf
import numpy as np
import pickle

def weight_and_bias(name ,shape):
    weight = tf.get_variable("W" + name, shape=shape, initializer=tf.contrib.layers.xavier_initializer())
    bias = tf.get_variable("B" + name, shape=shape[-1], initializer=tf.contrib.layers.xavier_initializer())
    return weight, bias

def conv2d_2x2(x, W):
    return tf.nn.conv2d(x, W, strides=[1, 5, 5, 1], padding='SAME')

def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

sess = tf.InteractiveSession()

source = tf.placeholder(tf.float32, [None, None, 50, 50])
source_len = tf.placeholder(tf.int32, [None])
source_max_step = tf.shape(source)[1]

target = tf.placeholder(tf.float32, [None, None, 50, 50])
target_len = tf.placeholder(tf.int32, [None])
target_max_step = tf.shape(target)[1]

W_conv, B_conv = weight_and_bias('conv1', [5, 5, 1, 32])

source = tf.reshape(source, [-1, 50, 50], "source_reshape")
source_tmp = tf.reshape(source, [-1, 50, 50 ,1])
source_conv = tf.nn.relu(conv2d_2x2(source_tmp, W_conv) + B_conv)
source_pool = max_pool_2x2(source_conv)
source_flat = tf.reshape(source_pool, [-1, 5 * 5 * 32], "source_pool_reshape")
source = tf.reshape(source_flat, [-1, source_max_step, 5*5*32], "source_flat_reshape")

W_conv, B_conv = weight_and_bias('conv2', [5, 5, 1, 32])

target = tf.reshape(target, [-1, 50, 50], "target_reshape")
target_tmp = tf.reshape(target, [-1, 50, 50 ,1])
target_conv = tf.nn.relu(conv2d_2x2(target_tmp, W_conv) + B_conv)
target_pool = max_pool_2x2(target_conv)
target_flat = tf.reshape(target_pool, [-1, 5 * 5 * 32], "target_pool_reshape")
target = tf.reshape(target_flat, [-1, target_max_step, 5*5*32], "target_flat_reshape")

source_cell = tf.nn.rnn_cell.LSTMCell(500, initializer=tf.contrib.layers.xavier_initializer())
target_cell = tf.nn.rnn_cell.LSTMCell(500, initializer=tf.contrib.layers.xavier_initializer())
source_rnn_output, _ = tf.nn.dynamic_rnn(source_cell, source, source_len, dtype=tf.float32, scope = "source")
target_rnn_output, _ = tf.nn.dynamic_rnn(target_cell, target, target_len, dtype=tf.float32, scope = "target")

source_output = tf.transpose(source_rnn_output, [1, 0, 2])
target_output = tf.transpose(target_rnn_output, [1, 0, 2])

source_final_output = tf.gather(source_output, -1)
target_final_output = tf.gather(target_output, -1)

output = tf.concat(1, [source_final_output, target_final_output])

W_sf, B_sf = weight_and_bias('sf', [1000, 2])
predict = tf.nn.softmax(tf.matmul(output, W_sf) + B_sf)
y = tf.placeholder(tf.float32, [None, 2])

cross_entropy = -tf.reduce_sum(y * tf.log(predict))
train_step = tf.train.RMSPropOptimizer(1e-4).minimize(cross_entropy)

correct_prediction = tf.equal(tf.arg_max(predict, 1), tf.arg_max(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

with open('set', 'rb') as f:
    _set = pickle.load(f)
    training_set = _set[0]
    training_len = _set[1]
    training_label = _set[2]

sess.run(tf.global_variables_initializer())

for i in range(20000):

    if i % 100 == 0:
        train_accuacy = accuracy.eval(feed_dict = {source: training_set[0], target: training_set[1], source_len: training_len[0], target_len: training_len[1], y: training_label})
        print("step %d, training accuracy %g"%(i, train_accuacy))
    train_step.run(feed_dict = {source: training_set[0], target: training_set[1], source_len: training_len[0], target_len: training_len[1], y: training_label})

这些是我的全部代码，我在其中找不到任何问题。

但是ValueError: Cannot feed value of shape (1077, 27, 50, 50) for Tensor 'source_flat_reshape:0', which has shape '(?, ?, 800)'被提出了。

错误消息很奇怪，因为它似乎发生在source = tf.reshape(source_flat, [-1, source_max_step, 5*5*32], "source_flat_reshape")，但source_flat的形状如何(1077, 27, 50, 50)？它应该是(1077*77, 800)

而且，有时会提出另一个ValueError: Cannot feed value of shape (1077, 27, 50, 50) for Tensor 'Reshape:0', which has shape '(?, 50, 50)'。

这也很难理解，为什么会发生？

希望有人能帮助我。

1 个答案:

答案 0 :(得分：1)

查看使用feed_dict时会发生什么 - 您引用变量source和target。但是，python变量不再引用占位符而是重构操作 - 因此op被“跳过”。

最简单的解决方法是将占位符重命名为独特的内容。进一步在网络中可以重复使用相同的名称（你可以只调用每一层net），只要你不再需要引用它们就无所谓了。

试试这个吗？

import tensorflow as tf
import numpy as np
import pickle

def weight_and_bias(name ,shape):
    weight = tf.get_variable("W" + name, shape=shape, initializer=tf.contrib.layers.xavier_initializer())
    bias = tf.get_variable("B" + name, shape=shape[-1], initializer=tf.contrib.layers.xavier_initializer())
    return weight, bias

def conv2d_2x2(x, W):
    return tf.nn.conv2d(x, W, strides=[1, 5, 5, 1], padding='SAME')

def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

sess = tf.InteractiveSession()

source_placeholder = tf.placeholder(tf.float32, [None, None, 50, 50])
source_len = tf.placeholder(tf.int32, [None])
source_max_step = tf.shape(source)[1]

target_placeholder = tf.placeholder(tf.float32, [None, None, 50, 50])
target_len = tf.placeholder(tf.int32, [None])
target_max_step = tf.shape(target)[1]

W_conv, B_conv = weight_and_bias('conv1', [5, 5, 1, 32])

source = tf.reshape(source_placeholder, [-1, 50, 50], "source_reshape")
source_tmp = tf.reshape(source, [-1, 50, 50 ,1])
source_conv = tf.nn.relu(conv2d_2x2(source_tmp, W_conv) + B_conv)
source_pool = max_pool_2x2(source_conv)
source_flat = tf.reshape(source_pool, [-1, 5 * 5 * 32], "source_pool_reshape")
source = tf.reshape(source_flat, [-1, source_max_step, 5*5*32], "source_flat_reshape")

W_conv, B_conv = weight_and_bias('conv2', [5, 5, 1, 32])

target = tf.reshape(target_placeholder, [-1, 50, 50], "target_reshape")
target_tmp = tf.reshape(target, [-1, 50, 50 ,1])
target_conv = tf.nn.relu(conv2d_2x2(target_tmp, W_conv) + B_conv)
target_pool = max_pool_2x2(target_conv)
target_flat = tf.reshape(target_pool, [-1, 5 * 5 * 32], "target_pool_reshape")
target = tf.reshape(target_flat, [-1, target_max_step, 5*5*32], "target_flat_reshape")

source_cell = tf.nn.rnn_cell.LSTMCell(500, initializer=tf.contrib.layers.xavier_initializer())
target_cell = tf.nn.rnn_cell.LSTMCell(500, initializer=tf.contrib.layers.xavier_initializer())
source_rnn_output, _ = tf.nn.dynamic_rnn(source_cell, source, source_len, dtype=tf.float32, scope = "source")
target_rnn_output, _ = tf.nn.dynamic_rnn(target_cell, target, target_len, dtype=tf.float32, scope = "target")

source_output = tf.transpose(source_rnn_output, [1, 0, 2])
target_output = tf.transpose(target_rnn_output, [1, 0, 2])

source_final_output = tf.gather(source_output, -1)
target_final_output = tf.gather(target_output, -1)

output = tf.concat(1, [source_final_output, target_final_output])

W_sf, B_sf = weight_and_bias('sf', [1000, 2])
predict = tf.nn.softmax(tf.matmul(output, W_sf) + B_sf)
y = tf.placeholder(tf.float32, [None, 2])

cross_entropy = -tf.reduce_sum(y * tf.log(predict))
train_step = tf.train.RMSPropOptimizer(1e-4).minimize(cross_entropy)

correct_prediction = tf.equal(tf.arg_max(predict, 1), tf.arg_max(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

with open('set', 'rb') as f:
    _set = pickle.load(f)
    training_set = _set[0]
    training_len = _set[1]
    training_label = _set[2]

sess.run(tf.global_variables_initializer())

for i in range(20000):

    if i % 100 == 0:
        train_accuacy = accuracy.eval(feed_dict = {source_placeholder: training_set[0], target_placeholder: training_set[1], source_len: training_len[0], target_len: training_len[1], y: training_label})
        print("step %d, training accuracy %g"%(i, train_accuacy))
    train_step.run(feed_dict = {source_placeholder: training_set[0], target_placeholder: training_set[1], source_len: training_len[0], target_len: training_len[1], y: training_label})
相关问题
最新问题