时间戳

首页 > 解决方案 > Tensorflow CNN 零精度？

python-3.x - Tensorflow CNN 零精度？

问题描述

我有一个包含 25000 张彩色图片 100*100(*3) 的数据集，我正在尝试用一个卷积层构建一个简单的神经网络。它的细胞图片是否被疟疾感染，所以我的输出是 2。但是对于每批，我得到 0% 的准确度。我的批次大小为 1，但我尝试了其他大小，但准确度仍然为 0%。

我的有线电视新闻网：

def simple_nn(X_training, Y_training, X_test, Y_test):
    input = 100*100*3
    h1 = 100
    batch_size = 1
    learning_rate = 0.000001
    dropout = 0.2

    X = tf.placeholder(tf.float32, [batch_size, 100, 100, 3], name="is_train")
    Y_ = tf.placeholder(tf.float32, [None, 2])

    #Layers
    conv1 = tf.layers.conv2d(X, filters=64, kernel_size=4,
                         strides=2, padding='SAME',
                         activation=tf.nn.relu, name="conv1")
    conv1 = tf.layers.batch_normalization(conv1)
    conv1 = tf.layers.max_pooling2d(conv1, 2, 2)

    conv2 = tf.layers.conv2d(conv1, filters=128, kernel_size=3,
                         strides=2, padding='SAME',
                         activation=tf.nn.relu, name="conv2")
    conv2 = tf.layers.dropout(conv2, rate=dropout)

    conv3 = tf.layers.conv2d(conv2, filters=256, kernel_size=3,
                     strides=2, padding='SAME',
                     activation=tf.nn.relu, name="conv3")
    conv3 = tf.layers.dropout(conv3, rate=dropout)

    conv4 = tf.layers.conv2d(conv3, filters=64, kernel_size=3,
                     strides=2, padding='SAME',
                     activation=tf.nn.relu, name="conv4")
    conv4 = tf.layers.max_pooling2d(conv4, 2, 2)

    conv5 = tf.layers.conv2d(conv4, filters=32, kernel_size=3,
                         strides=2, padding='SAME',
                         activation=tf.nn.relu, name="conv5")
    Y = tf.reshape(conv5, [batch_size,-1])
    logits = tf.layers.dense(Y, units=2, activation=tf.nn.relu)

    # loss function
    cross_entropy = tf.nn.softmax_cross_entropy_with_logits_v2(labels=Y_, logits=logits)
    loss = tf.reduce_mean(tf.cast(cross_entropy, tf.float32))

    # % of correct answers found in batch
    is_correct = tf.equal(tf.argmax(Y,1), tf.argmax(Y_,1))
    accuracy = tf.reduce_mean(tf.cast(is_correct, tf.float32))


    optimizer = tf.train.GradientDescentOptimizer(learning_rate)
    train_step = optimizer.minimize(cross_entropy)

    init = tf.global_variables_initializer()

    sess = tf.Session()
    sess.run(init)

    for i in range(math.floor(len(X_training)/batch_size)):
        st = batch_size * i
        end = st + batch_size

        if end >= math.floor(len(X_training)) - batch_size:
            break
        batch_X, batch_Y = X_training[st:end], Y_training[st:end]
        train_data={X: batch_X, Y_: batch_Y}

        sess.run(train_step, feed_dict=train_data)

        #Get the accuracy and loss
        a, l = sess.run([accuracy, cross_entropy], feed_dict=train_data)
        print("acc : "+str(a)+" , loss : "+str(l))

我的输出：

acc : 0.0 , loss : [0.6931472]
acc : 0.0 , loss : [0.6931472]
acc : 0.0 , loss : [0.69436306]
acc : 0.0 , loss : [0.6931472]
acc : 0.0 , loss : [0.6931662]
acc : 0.0 , loss : [0.6931472]
acc : 0.0 , loss : [0.6925567]
acc : 0.0 , loss : [0.6931472]
acc : 0.0 , loss : [0.6931472]
acc : 0.0 , loss : [0.69259375]
acc : 0.0 , loss : [0.6912933]
acc : 0.0 , loss : [0.6957785]
acc : 0.0 , loss : [0.6931472]
acc : 0.0 , loss : [0.6990725]
acc : 0.0 , loss : [0.69037354]
acc : 0.0 , loss : [0.6931472]
acc : 0.0 , loss : [0.6931472]
acc : 0.0 , loss : [0.6931472]
acc : 0.0 , loss : [0.6931472]
acc : 0.0 , loss : [0.6991633]
acc : 0.0 , loss : [0.6931472]
acc : 0.0 , loss : [0.6931472]
acc : 0.0 , loss : [0.6931472]
acc : 0.0 , loss : [0.700589]
acc : 0.0 , loss : [0.6931472]

我用一个简单的非卷积层得到了 65%（意思是acc=0.65），但是自从我切换到 conv 之后，acc=0.0. 首先，由于某种原因，我loss在使用卷积层时在我的变量中返回了准确性，但我现在不这么认为，我认为我的损失函数有问题。即使我将模型减少到一层，同样的事情也会发生，而且我loss的模型仍然存在0.69。

标签： python-3.xtensorflowdeep-learningconv-neural-network

解决方案

您应该最小化减少的向量。更改此行

train_step = optimizer.minimize(cross_entropy)

对此：

train_step = optimizer.minimize(loss)

此外，您不包括logits在准确性计算中的层。做这个：

is_correct = tf.equal(tf.argmax(logits,1), tf.argmax(Y_,1))
accuracy = tf.reduce_mean(tf.cast(is_correct, tf.float32))

此外，您正在对logits图层应用两个激活。首先你有tf.nn.relu然后你使用softmax(with tf.nn.softmax_cross_entropy_with_logits_v2())。不确定你是故意这样做的。

推荐阅读