我目前正在Tensorflow中开发一个程序,该程序读取1750 x 1750像素的数据。我通过一个卷积网络运行它:

import os
import sys

import tensorflow as tf
import Input

FLAGS = tf.app.flags.FLAGS

tf.app.flags.DEFINE_integer('batch_size', 100, "hello")
tf.app.flags.DEFINE_string('data_dir',     '/Volumes/Machine_Learning_Data',  "hello")

def inputs():
  if not FLAGS.data_dir:
    raise ValueError('Please supply a data_dir')
  data_dir = os.path.join(FLAGS.data_dir, 'Data')
  images, labels = Input.inputs(data_dir = data_dir, batch_size =     FLAGS.batch_size)
  return images, labels

def weight_variable(shape):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial)

def bias_variable(shape):
    initial = tf.constant(0.1, shape = shape)
    return tf.Variable(initial)

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

def max_pool_5x5(images):
    return tf.nn.max_pool(images, ksize = [1, 5, 5, 1], strides = [1, 1, 1, 1], padding = 'SAME')

def forward_propagation(images):
  with tf.variable_scope('conv1') as scope:
      W_conv1 = weight_variable([5, 5, 1, 32])
      b_conv1 = bias_variable([32])
      image_matrix = tf.reshape(images, [-1, 1750, 1750, 1])
      h_conv1 = tf.nn.sigmoid(conv2d(image_matrix, W_conv1) + b_conv1)
      h_pool1 = max_pool_5x5(h_conv1)

  with tf.variable_scope('conv2') as scope:
      W_conv2 = weight_variable([5, 5, 32, 64])
      b_conv2 = bias_variable([64])
      h_conv2 = tf.nn.sigmoid(conv2d(h_pool1, W_conv2) + b_conv2)
      h_pool2 = max_pool_5x5(h_conv2)

  with tf.variable_scope('conv3') as scope:
      W_conv3 = weight_variable([5, 5, 64, 128])
      b_conv3 = bias_variable([128])
      h_conv3 = tf.nn.sigmoid(conv2d(h_pool2, W_conv3) + b_conv3)
      h_pool3 = max_pool_5x5(h_conv3)

  with tf.variable_scope('local3') as scope:
      W_fc1 = weight_variable([10 * 10 * 128, 256])
      b_fc1 = bias_variable([256])
      h_pool3_flat = tf.reshape(h_pool3, [-1, 10 * 10 * 128])
      h_fc1 = tf.nn.sigmoid(tf.matmul(h_pool3_flat, W_fc1) + b_fc1)
      keep_prob = tf.placeholder(tf.float32)
      h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
      W_fc2 = weight_variable([256, 4])
      b_fc2 = bias_variable([4])

      y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
      return y_conv

def error(forward_propagation_results, labels):
    labels = tf.cast(labels, tf.float32)
    mean_squared_error = tf.square(tf.sub(labels, forward_propagation_results))
    cost = tf.reduce_mean(mean_squared_error)
    train = tf.train.GradientDescentOptimizer(learning_rate = 0.3).minimize(cost)
    return train

print cost


不幸的是弹出了一个错误

用于广播的不兼容形状:TensorShape([Dimension(100)])和TensorShape([Dimension(9187500),Dimension(4)])

而且我无法调试它。

矩阵尺寸有什么问题?解释器说错误发生在tf.sub行。

编辑:

这是调用函数的代码的主要部分。

import Input
import Process

import tensorflow as tf


def train():
    with tf.Session() as sess:
        images, labels = Process.inputs()

        forward_propgation_results = Process.forward_propagation(images)

        train_loss = Process.error(forward_propgation_results, labels)

        init = tf.initialize_all_variables()

        sess.run(init)

def main(argv = None):
    train()

if __name__ == '__main__':
  tf.app.run()

最佳答案

我发现了以下问题:


您的labels输入是一个简单的一维标签标识符数组,但需要对其进行一热编码,以使其成为尺寸为[batch_size, 4]且填充为1或0的矩阵。
您的最大池化操作需要具有不同于1的跨度,以实际减小图像的宽度和高度。因此,设置strides=[1, 5, 5, 1]应该可以。
解决此问题后,您的最大池化操作实际上并没有像您假设的那样将宽度/高度从1750降低到10,而是降低到14(因为1750 / 5 / 5 / 5 == 14。所以您可能想在这里增加权重矩阵,但是还有其他选择。
您的图像是否有可能以3个频道开头?您在这里假设为灰度,因此您应该重塑image_matrix使其具有3个通道,或者将图像转换为灰度。


应用这些修复程序之后,网络输出和标签都应具有形状[batch_size, 4],并且您应该能够计算出差异。

编辑:在下面的聊天中讨论了代码之后,我已经对此进行了调整。

09-11 19:47