## 读取数据集X, Y = read_dataset()model_path = "/Users/shalinsavalia/Desktop/ECG_CNN/CNN" ## 打乱数据集以混合行X, Y = shuffle(X, Y, random_state=1) ## 将数据集分为训练和测试部分train_x, test_x, train_y, test_y = train_test_split(X, Y, test_size=0.20, random_state=415)print "训练数据集中ECG信号的形状 {}".format(train_x.shape)print "训练数据集中ECG疾病的形状 {}".format(train_y.shape)print "测试数据集中ECG信号的形状 {}".format(test_x.shape)print "测试数据集中ECG疾病的形状 {}".format(test_y.shape) ## 定义处理张量所需的重要参数和变量n_dim = X.shape[1]n_class = 9 # 输入 - 形状中的'None'表示可以是任意值，即我们可以输入任意数量的ECG信号 ## 输入ECG信号长度x=tf.placeholder(tf.float32,shape=[None,n_dim]) ## 输入类别（共有9种疾病）y_=tf.placeholder(tf.float32,shape=[None, n_class]) ## 重塑输入信号以便输入到CNNx_input=tf.reshape(x,[-1,1,60,1], name='input') ## 第一个卷积层，输出滤波器32个，滤波器大小5x5，步长为2，同步填充，RELU激活函数conv_layer1=tflearn.layers.conv.conv_2d(x_input, nb_filter=32, filter_size=5, strides=[1,1,1,1], padding='same', activation='relu', regularizer="L2", name='conv_layer_1')out_layer1=tflearn.layers.conv.max_pool_2d(conv_layer1, 2) ## 第二个卷积层conv_layer2=tflearn.layers.conv.conv_2d(out_layer1, nb_filter=128, filter_size=5, strides=[1,1,1,1], padding='same', activation='relu', regularizer="L2", name='conv_layer_2')out_layer2=tflearn.layers.conv.max_pool_2d(conv_layer2, 2) ## 第三个卷积层 conv_layer3=tflearn.layers.conv.conv_2d(out_layer2, nb_filter=64, filter_size=5, strides=[1,1,1,1], padding='same', activation='relu', regularizer="L2", name='conv_layer_3')out_layer3=tflearn.layers.conv.max_pool_2d(conv_layer3, 2) ## 第四个卷积层 conv_layer4=tflearn.layers.conv.conv_2d(out_layer3, nb_filter=32, filter_size=5, strides=[1,1,1,1], padding='same', activation='relu', regularizer="L2", name='conv_layer_4')out_layer4=tflearn.layers.conv.max_pool_2d(conv_layer4, 2) ## 全连接层fcl= tflearn.layers.core.fully_connected(out_layer4, 1024, activation='relu')fcl_dropout = tflearn.layers.core.dropout(fcl, 0.8)y_predicted = tflearn.layers.core.fully_connected(fcl_dropout, 9, activation='softmax', name='output') ## 损失函数cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y_predicted), reduction_indices=[1])) ## 优化器 train_step = tf.train.GradientDescentOptimizer(0.2).minimize(cross_entropy) ## 计算模型的准确率 correct_prediction = tf.equal(tf.argmax(y_predicted,1), tf.argmax(y_,1))accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))结果看起来像这样
测试数据集中ECG疾病的形状 (5309, 9)输入的形状 : [None, 1, 60, 1]第一个卷积层的形状 : [None, 1, 30, 32]第二个卷积层的形状 : [None, 1, 15, 128]第三个卷积层的形状 : [None, 1, 15, 64]第四个卷积层的形状 : [None, 1, 15, 32]全连接层的形状 : [None, 1024]输出层的形状 : [None, 9]('Epoch :', 198, '- Loss: ', 1.046878, '- Train Accuracy: ',0.6050674)('Epoch :', 199, '- Loss: ', 1.0282028, '- Train Accuracy: ',0.60676277)('Epoch :', 200, '- Loss: ', 1.0455937, '- Train Accuracy: ',0.59508336)('Epoch :', 201, '- Loss: ', 1.0259817, '- Train Accuracy: ',0.6227277)('Epoch :', 202, '- Loss: ', 1.0068184, '- Train Accuracy: ',0.61241406)('Epoch :', 203, '- Loss: ', 0.9958509, '- Train Accuracy: ',0.6385985)('Epoch :', 204, '- Loss: ', nan, '- Train Accuracy: ', 0.6148159)('Epoch :', 205, '- Loss: ', nan, '- Train Accuracy: ', 0.118065365)('Epoch :', 206, '- Loss: ', nan, '- Train Accuracy: ', 0.118065365)('Epoch :', 207, '- Loss: ', nan, '- Train Accuracy: ', 0.118065365)('Epoch :', 208, '- Loss: ', nan, '- Train Accuracy: ', 0.118065365)从我的训练结果来看，损失值最初是在下降，直到第203个epoch，突然就变成了nan。我尝试更改卷积层的数量、最大池化层，但仍然没有效果。在所有情况下，经过几个epoch后损失值都会变成nan。我观察到，在所有情况下，当损失值达到大约0.98或0.99时，之后就会变成nan。我不知道我的代码哪里出了问题？请帮帮我，我会非常感激的。
回答：
以下交叉熵代码不安全
cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y_predicted), reduction_indices=[1]))因为当y_predicted传递零值给log时，log的值是未定义的。
要解决这个问题，请使用以下代码
cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(tf.clip_by_value(y_predicted,1e-10,1.0)), reduction_indices=[1]))