如何使用 shuffle_batch() 函数避免超出范围错误？

Question

更新问题

我正在尝试使用shuffle_batch() 函数将标签与张量流中的图像进行匹配，但是当我使用 shuffle_batch() 函数开始训练循环时，出现 RandomShuffleQueue 超出范围错误。

1。我更新的问题

使用shuffle_batch函数时如何避免队列超出范围错误？

2.我的更新代码

以下代码在前 90 步左右运行良好，并且精度越来越高，直到它引发错误。

# Global Parameters

# Image Size
training_size = 1387
img_height = 64
img_width = 64

# File stream
batch_size = 128

# Training parameter
learning_rate = 0.001
training_iters = 100
keep_prob = 0.5 #dropout keep prob
display_step = 10

AdamOptimizer = 1
GradientDescentOptimizer = 0

# Filepath
csv_filepath = r'C:/Users/Jeffy/OneDrive/Course\NMDA\retinaProject\label.csv'
image_filepath = 'Image_P/'


import tensorflow as tf
# =============================================================================
# Read input data

# load csv content
csv_path = tf.train.string_input_producer(['label_3D.csv'])
textReader = tf.TextLineReader()
_, csv_content = textReader.read(csv_path)
im_name, col_2, col_3, col_4 = tf.decode_csv(csv_content, record_defaults=[[""], [1], [1], [1]])
label = tf.pack([col_2, col_3, col_4])
label_float32 = tf.cast(label, tf.float32)

# load images
im_content = tf.read_file(image_filepath + im_name+'.jpeg')
image = tf.image.decode_jpeg(im_content, channels=3)
image_float32 = tf.cast(image, tf.float32)/255

# Generate Batch
batch_shape = ((img_height, img_width, 3),(3))
images_batch, labels_batch = tf.train.shuffle_batch([image_float32, label_float32], 
                                                    batch_size = batch_size, 
                                                    capacity = batch_size * 50, 
                                                    min_after_dequeue = batch_size * 10, 
                                                    shapes = batch_shape)

# =============================================================================
# Construct Network
# define functions
def weight_varible(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(x, W):
    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 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')


# paras
W_conv1 = weight_varible([5, 5, 3, 32])
b_conv1 = bias_variable([32])

# conv layer-1
h_conv1 = tf.nn.relu(conv2d(images_batch, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)

# conv layer-2
W_conv2 = weight_varible([5, 5, 32, 64])
b_conv2 = bias_variable([64])

h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)

# full connection
W_fc1 = weight_varible([16 * 16 * 64, 1024])
b_fc1 = bias_variable([1024])

h_pool2_flat = tf.reshape(h_pool2, [-1, 16 * 16 * 64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

# dropout

h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

# output layer: softmax
W_fc2 = weight_varible([1024, 3])
b_fc2 = bias_variable([3])


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

# model training
cross_entropy = -tf.reduce_sum(labels_batch * tf.log(y_conv))

if GradientDescentOptimizer:
    train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(cross_entropy)
else:
    if AdamOptimizer:
        train_step = tf.train.AdamOptimizer(learning_rate).minimize(cross_entropy)

correct_prediction = tf.equal(tf.arg_max(y_conv, 1), tf.arg_max(labels_batch, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))


with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    # Start file queue
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(coord=coord)

    sess.run([images_batch, labels_batch])

    coord.request_stop()
    coord.join(threads)

    for i in range(training_iters):
        # display the result on console
        if i % display_step == 0:
            train_accuacy = accuracy.eval()
            print("step %d, training accuracy %g"%(i, train_accuacy))
        # run the model
        train_step.run()
    print("test accuracy %g"%(accuracy.eval()))

3.更新运行结果

更新后的代码在第 90 步左右调用错误：

step 0, training accuracy 0.5625
step 10, training accuracy 0.6875
step 20, training accuracy 0.703125
step 30, training accuracy 0.625
step 40, training accuracy 0.65625
step 50, training accuracy 0.6875
step 60, training accuracy 0.6875
step 70, training accuracy 0.734375
step 80, training accuracy 0.632812
step 90, training accuracy 0.695312

然后

OutOfRangeError: RandomShuffleQueue '_24_shuffle_batch_3/random_shuffle_queue' is closed and has insufficient elements (requested 128, current size 1)
     [[Node: shuffle_batch_3 = QueueDequeueMany[_class=["loc:@shuffle_batch_3/random_shuffle_queue"], component_types=[DT_FLOAT, DT_FLOAT], timeout_ms=-1, _device="/job:localhost/replica:0/task:0/cpu:0"](shuffle_batch_3/random_shuffle_queue, shuffle_batch_3/n)]]

Caused by op 'shuffle_batch_3', defined at:
  File "C:\Users\Jeffy\Anaconda3\lib\site-packages\spyder\utils\ipython\start_kernel.py", line 223, in <module>
    main()
......

---

老问题：

为什么当我使用 tf.batch() 将标签与图像匹配时，图像的标签会粘在第一个样本上？ 在 Standy 和我的朋友 Dong 的大力帮助下，我的老问题得到了解决。非常感谢他们！

1.我的问题

1. 将“shuffle_batch()”函数放在位置 1 或位置 2 的什么位置？ （你可以在下面代码的 tf.session() 部分找到它）

2. 如何修复代码以使批量成功组合图像和标签，而不是卡在第一个示例上？

2.csv 文件

图像名称和相关标签以以下格式（不带标题）存储在 label.csv 中：

11219_right,0,1,0
15502_left,0,0,0
14481_right,0,1,0
11032_right,0,0,1
19322_right,0,0,0
......

3.源代码

代码的目的是使用RNN作为图像分类器。

CNN结构基于tensorflow示例文件。

您可以专注于读取数据部分并启动图形部分。

# Global Parameters

# Image Size
training_size = 1387
img_height = 64
img_width = 64

# File stream
batch_size = 128

# Training parameter
learning_rate = 0.001
training_iters = 100
keep_prob = 0.5 #dropout keep prob
display_step = 10

AdamOptimizer = 1
GradientDescentOptimizer = 0

# Filepath
csv_filepath = r'C:/Users/Jeffy/OneDrive/Course\NMDA\retinaProject\label.csv'
image_filepath = 'Image_P/'


# import library
import tensorflow as tf
import numpy as np
#=============================================================================
# Read input data

# load csv content
csv_path = tf.train.string_input_producer(['label.csv'])
textReader = tf.TextLineReader()
_, csv_content = textReader.read(csv_path)
im_name, label = tf.decode_csv(csv_content, record_defaults=[[""], [1]])

# load images
im_content = tf.read_file(image_filepath + im_name+'.jpeg')
image = tf.image.decode_jpeg(im_content, channels=3)

def label_3D (label_num):
    label_3D = np.zeros(3)
    if label_num == 0:
        label_3D[0] = 1
    else:
        if label_num == 3:
            label_3D[1] = 1
        else: # label_num == 4
            label_3D[2] = 1
    return label_3D
# =============================================================================
# Construct Network(you can skip this part)

# define functions
def weight_varible(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(x, W):
    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 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')


# paras
W_conv1 = weight_varible([5, 5, 3, 32])
b_conv1 = bias_variable([32])

# conv layer-1
x = tf.Variable(tf.zeros([batch_size, img_width, img_height, 3]))
h_conv1 = tf.nn.relu(conv2d(x, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)

# conv layer-2
W_conv2 = weight_varible([5, 5, 32, 64])
b_conv2 = bias_variable([64])

h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)

# full connection
W_fc1 = weight_varible([16 * 16 * 64, 1024])
b_fc1 = bias_variable([1024])

h_pool2_flat = tf.reshape(h_pool2, [-1, 16 * 16 * 64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

# dropout

h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

# output layer: softmax
W_fc2 = weight_varible([1024, 3])
b_fc2 = bias_variable([3])


y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)
y_ = tf.Variable(tf.zeros([batch_size, 3]))

# model training
cross_entropy = -tf.reduce_sum(y_ * tf.log(y_conv))
if GradientDescentOptimizer:
    train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(cross_entropy)
else:
    if AdamOptimizer:
        train_step = tf.train.AdamOptimizer(learning_rate).minimize(cross_entropy)

correct_prediction = tf.equal(tf.arg_max(y_conv, 1), tf.arg_max(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

# ==========================================================================
# Lauch the graph
with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    # Start file queue
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(coord=coord)

    images, labels = sess.run([image, label])
    #position1 for tf.train.shuffle_batch() function

    coord.request_stop()
    coord.join(threads)

    for i in range(training_iters):
        #position2 for tf.train.shuffle_batch() function
        batch = tf.train.shuffle_batch([images,label_3D(labels)], batch_size=batch_size,
                               capacity = batch_size * 50, 
                               min_after_dequeue = batch_size * 10,
                               num_threads = 1)

        if i % display_step == 0:
            x = batch[0]
            y_ = batch[1]
            train_accuacy = accuracy.eval()
            print("step %d, training accuracy %g"%(i, train_accuacy))
        x= batch[0]
        y_ = batch[1]
        train_step.run()

`

4.运行结果 - 在 IPython 控制台中
step 0, training accuracy 0.226562 step 10, training accuracy 1 step 20, training accuracy 1 step 30, training accuracy 1 step 40, training accuracy 1 - 在变量探索中

变量labels和images和label和image first example一样，保存在label.csv文件的第一行

因此，我推断读取文件队列卡在了第一行，导致CNN快速收敛到1的精度。

Answer 1

shuffle_batch 接受张量并返回一个张量，所以它是一个张量流操作，应该放在Graph中。 shuffle_batch docs

我会在你解码单个图像后立即放置它：

image = tf.image.decode_jpeg(im_content, channels=3)
images_batch, labels_batch = tf.train.shuffle_batch([image, label], batch_size, batch_size * 50, batch_size * 10)
# images_batch is now Tensor of shape (batch_size, height, weight, channels)
...
h_conv1 = tf.nn.relu(conv2d(images_batch, W_conv1) + b_conv1)

您现在不需要变量x 和y_，并且在使用tf.train.shuffle_batch 时不需要手动分配输入。

tf.train.shuffle_batch 接受单个示例的张量并生成整个批次似乎违反直觉，但请记住，您提供给此操作的张量来自队列，因此 tf.train.shuffle_batch 可以“等待”多个元素（实际上它使用另一个队列来进行洗牌和存储中间元素，shuffle_batch 实现是here）