在了解Tensorflow的过程中，我构建了一个用于分类的玩具网络。该网络有15个输入节点，这些节点的特征与相应类别标签的一热编码相同（索引从1开始）。因此，从输入CSV文件加载的数据可能看起来像这样：

1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,10,1,0,0,0,0,0,0,0,0,0,0,0,0,0,2...0,0,0,0,0,0,0,0,0,0,0,0,0,0,1,15

该网络只有一个隐藏层和一个输出层，后者包含给定类的概率。我的问题是：在训练过程中，网络对作为第一个输入的任何内容分配了越来越高的概率。

以下是相关的代码行（省略了一些行）：

# number_of_p : 样本数# number_of_a : 属性数（特征） -> 15# number_of_s : 样式数（标签） -> 15# 生成隐藏层的函数# nodes是每层节点的列表（len(nodes) = 隐藏层的数量）def hidden_generation(nodes):    hidden_nodes = [number_of_a] + nodes + [number_of_s]    number_of_layers = len(hidden_nodes) - 1    print(hidden_nodes)    hidden_layer = list()    for i in range (0,number_of_layers):        hidden_layer.append(tf.zeros([hidden_nodes[i],batch_size]))    hidden_weights = list()    for i in range (0,number_of_layers):        hidden_weights.append(tf.Variable(tf.random_normal([hidden_nodes[i+1], hidden_nodes[i]])))    hidden_biases = list()    for i in range (0,number_of_layers):        hidden_biases.append(tf.Variable(tf.zeros([hidden_nodes[i+1],batch_size])))    return hidden_layer, hidden_weights, hidden_biases#损失函数def loss(labels, logits):    cross_entropy = tf.losses.softmax_cross_entropy(        onehot_labels = labels, logits = logits)    return tf.reduce_mean(cross_entropy, name = 'xentropy_mean')hidden_layer, hidden_weights, hidden_biases = hidden_generation(hidden_layers)with tf.Session() as training_sess:    training_sess.run(tf.global_variables_initializer())    training_sess.run(a_iterator.initializer, feed_dict = {a_placeholder_feed: training_set.data})    current_a = training_sess.run(next_a)       training_sess.run(s_iterator.initializer, feed_dict = {s_placeholder_feed: training_set.target})    current_s = training_sess.run(next_s)     s_one_hot = training_sess.run(tf.one_hot((current_s - 1), number_of_s))    for i in range (1,len(hidden_layers)+1):        hidden_layer[i] = tf.tanh(tf.matmul(hidden_weights[i-1], (hidden_layer[i-1])) + hidden_biases[i-1])    output = tf.nn.softmax(tf.transpose(tf.matmul(hidden_weights[-1],hidden_layer[-1]) + hidden_biases[-1]))    optimizer = tf.train.GradientDescentOptimizer(learning_rate = 0.1)    # 使用AdamOptimizer无济于事，选择更大或更小的学习率也无济于事    train = optimizer.minimize(loss(s_one_hot, output))    training_sess.run(train)    for i in range (0, (number_of_p)):        current_a = training_sess.run(next_a)        current_s = training_sess.run(next_s)        s_one_hot = training_sess.run(tf.transpose(tf.one_hot((current_s - 1), number_of_s)))        # （不知道为什么我必须声明两次才能移动数据流）        training_sess.run(train)

我认为损失函数的声明位置不正确，总是引用相同的向量。然而，替换损失函数对我目前没有帮助。如果有人愿意帮助我，我会很乐意提供剩余的代码。

编辑：我已经发现并修复了一个主要的（而且愚蠢的）错误：在tf.matmul中，权重应该放在节点值之前。

回答：

你不希望反复声明训练操作。这是不必要的，而且如你所指出的那样会更慢。你没有将current_a输入到神经网络中。所以你不会得到新的输出，你使用迭代器的方式也不正确，这也可能是问题的根源。

with tf.Session() as training_sess:    training_sess.run(tf.global_variables_initializer())    training_sess.run(a_iterator.initializer, feed_dict = {a_placeholder_feed: training_set.data})    current_a = training_sess.run(next_a)       training_sess.run(s_iterator.initializer, feed_dict = {s_placeholder_feed: training_set.target})    current_s = training_sess.run(next_s)     s_one_hot = training_sess.run(tf.one_hot((current_s - 1), number_of_s))    for i in range (1,len(hidden_layers)+1):        hidden_layer[i] = tf.tanh(tf.matmul(hidden_weights[i-1], (hidden_layer[i-1])) + hidden_biases[i-1])    output = tf.nn.softmax(tf.transpose(tf.matmul(hidden_weights[-1],hidden_layer[-1]) + hidden_biases[-1]))    optimizer = tf.train.GradientDescentOptimizer(learning_rate = 0.1)    # 使用AdamOptimizer无济于事，选择更大或更小的学习率也无济于事    train = optimizer.minimize(loss(s_one_hot, output))    training_sess.run(train)    for i in range (0, (number_of_p)):        current_a = training_sess.run(next_a)        current_s = training_sess.run(next_s)        s_one_hot = training_sess.run(tf.transpose(tf.one_hot((current_s - 1), number_of_s)))        # （不知道为什么我必须声明两次才能移动数据流）        training_sess.run(train)

这里有一些伪代码来帮助你获得正确的数据流。我建议在训练之前进行一热编码，这样在加载数据时会更容易处理。

train_dataset = tf.data.Dataset.from_tensor_slices((inputs, targets))train_dataset = train_dataset.batch(batch_size)train_dataset = train_dataset.repeat(num_epochs)iterator = train_dataset.make_one_shot_iterator()next_inputs, next_targets = iterator.get_next()# 定义训练过程global_step = tf.Variable(0, name="global_step", trainable=False)loss = Neural_net_function(next_inputs, next_targets)optimizer = tf.train.AdamOptimizer(learning_rate)grads_and_vars = optimizer.compute_gradients(loss)train_op = optimizer.apply_gradients(grads_and_vars, global_step=global_step)with tf.Session() as training_sess:    for i in range(number_of_training_samples * num_epochs):            taining_sess.run(train_op)