哪个参数决定输入张量的大小？

Question

在我的CNN网络中，我没有定义输入张量的大小，但是为什么网络如Tensorboard中所示有六个输入张量（参见图片）？

和Tensorboard直方图中的conv1_5，conv2_5，conv4_5，conv4_5，conv5_5，conv7_5，conv8_5 ，... ...最后一列的张量权重和偏差会逐步更改，而其他张量参数只是常量（步骤0的数据）。

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')

def weight_variable(shape, alias):
    initial = tf.truncated_normal(shape, stddev=0.1, name=alias)
    var = tf.Variable(initial)
    name_scope = tf.contrib.framework.get_name_scope()
    tf.summary.histogram(name_scope + '/weights', var)
    return var

def bias_variable(shape, alias):
    initial = tf.constant(0.1, shape=shape, name=alias)
    var = tf.Variable(initial)
    name_scope = tf.contrib.framework.get_name_scope()
    tf.summary.histogram(name_scope + '/bias', var)
    return tf.Variable(initial)

def cnn(x):
    #[n, 28, 28, 3]
    with tf.name_scope('reshape1'):
        x_image = tf.reshape(x, [-1, 28, 28, 3])

    #[n, 28, 28, 32]
    with tf.name_scope('conv1'):
        W_conv1 = weight_variable([3, 3, 3, 32], 'weight')
        b_conv1 = bias_variable([32], 'bias')
        layer_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)

     #[n, 28, 28, 32]
    with tf.name_scope('conv2'):
        W_conv2 = weight_variable([3, 3, 32, 32], 'weight')
        b_conv2 = bias_variable([32], 'bias')
        layer_conv2 = tf.nn.relu(conv2d(layer_conv1, W_conv2) + b_conv2)

    #[n, 14, 14, 32]
    with tf.name_scope('pool3'):
        layer_pool3 = max_pool_2x2(layer_conv2)

    #[n, 14, 14, 64]
    with tf.name_scope('conv4'):
        W_conv4 = weight_variable([3, 3, 32, 64], 'weight')
        b_conv4 = bias_variable([64], 'bias')
        layer_conv4 = tf.nn.relu(conv2d(layer_pool3, W_conv4) + b_conv4)

    #[n, 14, 14, 64]
    with tf.name_scope('conv5'):
        W_conv5 = weight_variable([3, 3, 64, 64], 'weight')
        b_conv5 = bias_variable([64], 'bias')
        layer_conv5 = tf.nn.relu(conv2d(layer_conv4, W_conv5) + b_conv5) 

    #[n, 7, 7, 64]
    with tf.name_scope('pool6'):
        layer_pool6 = max_pool_2x2(layer_conv5)

    #[n, 7, 7, 128]
    print(layer_pool6.shape)
    with tf.name_scope('conv7'):
        W_conv7 = weight_variable([3, 3, 64, 128], 'weight')
        b_conv7 = bias_variable([128], 'bias')
        layer_conv7 = tf.nn.relu(conv2d(layer_pool6, W_conv7) + b_conv7)

    #[n, 7, 7, 128]
    with tf.name_scope('conv8'):
        W_conv8 = weight_variable([3, 3, 128, 128], 'weight')
        b_conv8 = bias_variable([128], 'bias')
        layer_conv8 = tf.nn.relu(conv2d(layer_conv7, W_conv8) + b_conv8)

    #[n, 7, 7, 128]
    with tf.name_scope('reshape9'):
        layer_reshape9 = tf.reshape(layer_conv8, [-1, 7 * 7 * 128])

    #[n, 7, 7, 128]
    with tf.name_scope('fc10'):
        W_fc10 = weight_variable([7 * 7 * 128, 1024], 'weight')
        b_fc10 = bias_variable([1024], 'bias')
        layer_fc10 = tf.nn.relu(tf.matmul(layer_reshape9, W_fc10) + b_fc10)

    with tf.name_scope('dropout11'):
        dropout_prob = tf.placeholder(tf.float32)
        layer_dropout11 = tf.nn.dropout(layer_fc10, dropout_prob, name='prob')

    with tf.name_scope('fc12'):
        W_fc12 = weight_variable([1024, 10], 'weight')
        b_fc12 = bias_variable([10], 'bias')
        y = tf.matmul(layer_dropout11, W_fc12) + b_fc12

    #return y
    return y, dropout_prob


def next_batch_set(images, labels, batch_size=128):
    """Generate a batch training data.

    Args:
        images: A 4-D array representing the training images.
        labels: A 1-D array representing the classes of images.
        batch_size: An integer.

    Return:
        batch_images: A batch of images.
        batch_labels: A batch of labels.
    """
    indices = np.random.choice(len(images), batch_size)
    batch_images = images[indices]
    batch_labels = labels[indices]
    return batch_images, batch_labels


def captcha_cnn():

    x = tf.placeholder(tf.float32, shape=[None, 28, 28, 3], name='inputs')
    y_ = tf.placeholder(tf.int32, shape=[None], name='labels')

    x = preprocess(x)

    predict_labels, dropout_prob = cnn(x)
    with tf.name_scope('loss'):
        cross_entropy = tf.reduce_mean(
            tf.nn.sparse_softmax_cross_entropy_with_logits(
                logits=predict_labels, labels=y_))

    with tf.name_scope('adam_optimizer'):
        train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)

    with tf.name_scope('accuracy'):
        logits = tf.nn.softmax(predict_labels)
        classes = tf.cast(tf.argmax(logits, axis=1), dtype=tf.int32, name='predict')
        classes_ = tf.identity(classes, name='classes')
        accuracy = tf.reduce_mean(tf.cast(tf.equal(classes, y_), 'float'))

    train_images, train_targets = read_captcha_images(TRAIN_IMAGES_PATH)
    graph = tf.get_default_graph()
    conv1_var = graph.get_tensor_by_name('conv1/Variable:0')

    saver = tf.train.Saver()

    merged = tf.summary.merge_all()    

    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        #选定可视化存储目录    
        writer = tf.summary.FileWriter("./logs/",sess.graph)   

        for i in range(8000):
            batch_images, batch_labels = next_batch_set(train_images, train_targets)
            train_dict = {x: batch_images, y_: batch_labels, dropout_prob: 1}
            #train_dict = {x: batch_images, y_: batch_labels}
            sess.run(train_step, feed_dict=train_dict)

            if i % 100 == 0:
                loss_, acc_ , classes_= sess.run([cross_entropy, accuracy, classes], feed_dict=train_dict)
                train_text = 'step: {}, loss: {}, acc: {}, class: {}, labels: {}'.format(i, loss_, acc_, classes_, batch_labels)
                print(train_text)
                result = sess.run(merged, feed_dict=train_dict)
                writer.add_summary(result, i)

        test_images, test_targets = read_captcha_images(TEST_IMAGES_PATH)
        print('test accuracy %g' % accuracy.eval(feed_dict={
             x: test_images, y_: test_targets, dropout_prob: 1.0}))


        saver.save(sess, "./model/captcha_cnn_model_data")

        graph_def = graph.as_graph_def()
        output_graph_def = convert_variables_to_constants(sess, graph_def, output_node_names=['accuracy/predict'])
        with tf.gfile.FastGFile('./model/captcha_cnn_model.pb', mode='wb') as f:
            f.write(output_graph_def.SerializeToString())
            print(f)

Answer 1

我找到了原因，我使用jupyter笔记本编辑代码，然后运行了CNN单元六次。：( – –