多标签分类的 sigmoid 非线性阈值

我正在尝试使用DenseNet 架构对来自https://www.kaggle.com/nih-chest-xrays/data 的x 射线图像进行分类. 该模型生成一个二元标签向量，其中每个标签表示 14 种可能病理的存在或不存在：肺不张、心脏肥大、实变、水肿、积液、肺气肿、纤维化、疝气、浸润、肿块、结节、胸膜增厚、肺炎和气胸。例如，健康患者的标签为 [0,0,0,0,0,0,0,0,0,0,0,0,0,0]，而水肿和积液患者的标签为 [0,0,0,0,0,0,0,0,0,0,0,0,0,0] [0,0,0,1,1,0,0,0,0,0,0,0,0,0] 的标签。我用 tensorflow 构建了这个模型，因为这是一个多标签分类问题，所以我使用的成本函数是 tf.reduce_mean(tf.losses.sigmoid_cross_entropy(labels, logits))，它用 AdamOptimizer 最小化。但是，当我检查 sigmoid 输出时，这些值都低于 0.5，导致 tf.round(logits) 为每个预测生成零。不同输入的实际 logits 不同，并且在 10000 次迭代后为非零值，因此我认为梯度消失不是问题。我有两个问题：

1. 这个问题可能是由模型的错误实现引起的吗？
2. 如果我将 sigmoid 函数的阈值从 0.5 降低到 0.25 以提高模型精度，我会“作弊”吗？

谢谢。

这是模型的代码：

def DenseNet(features, labels, mode, params):

depth = params["depth"]
k = params["growth"]

if depth == 121:
    N = db_121
else:
    N = db_169

bottleneck_output = 4 * k

#before entering the first dense block, a conv operation with 16 output channels
#is performed on the input images

with tf.variable_scope('input_layer'):
    #l = tf.reshape(features, [-1, 224, 224, 1])
    feature_maps = 2 * k
    l = layers.conv(features, filter_size = 7, stride = 2, out_chn = feature_maps)
    l = tf.nn.max_pool(l,
                       padding='SAME',
                       ksize=[1,3,3,1],
                       strides=[1,2,2,1],
                       name='max_pool')

# each block is defined as a dense block + transition layer
with tf.variable_scope('block1'):
    for i in range(N[0]):
        with tf.variable_scope('bottleneck_layer.{}'.format(i+1)):
            bn_l = layers.batch_norm('BN', l)
            bn_l = tf.nn.relu(bn_l, name='relu')
            bn_l = layers.conv(bn_l, out_chn=bottleneck_output, filter_size=1)
        l = layers.add_layer('dense_layer.{}'.format(i+1), l, bn_l)
    l = layers.transition_layer('transition1', l)

with tf.variable_scope('block2'):
    for i in range(N[1]):
        with tf.variable_scope('bottleneck_layer.{}'.format(i+1)):
            bn_l = layers.batch_norm('BN', l)
            bn_l = tf.nn.relu(bn_l, name='relu')
            bn_l = layers.conv(bn_l, out_chn=bottleneck_output, filter_size=1)
        l = layers.add_layer('dense_layer.{}'.format(i+1), l, bn_l)
    l = layers.transition_layer('transition2', l)

with tf.variable_scope('block3'):
    for i in range(N[2]):
        with tf.variable_scope('bottleneck_layer.{}'.format(i+1)):
            bn_l = layers.batch_norm('BN', l)
            bn_l = tf.nn.relu(bn_l, name='relu')
            bn_l = layers.conv(bn_l, out_chn=bottleneck_output, filter_size=1)
        l = layers.add_layer('dense_layer.{}'.format(i+1), l, bn_l)
    l = layers.transition_layer('transition3', l)

# the last block does not have a transition layer
with tf.variable_scope('block4'):
    for i in range(N[3]):
        with tf.variable_scope('bottleneck_layer.{}'.format(i+1)):
            bn_l = layers.batch_norm('BN', l)
            bn_l = tf.nn.relu(bn_l, name='relu')
            bn_l = layers.conv(bn_l, out_chn=bottleneck_output, filter_size=1)
        l = layers.add_layer('dense_layer.{}'.format(i+1), l, bn_l)

# classification (global max pooling and softmax)
with tf.name_scope('classification'):
    l = layers.batch_norm('BN', l)
    l = tf.nn.relu(l, name='relu')
    l = layers.pooling(l, filter_size = 7)
    l_shape = l.get_shape().as_list()
    l = tf.reshape(l, [-1, l_shape[1] * l_shape[2] * l_shape[3]])
    l = tf.layers.dense(l, units = 1000, activation = tf.nn.relu, name='fc1', kernel_initializer=tf.contrib.layers.xavier_initializer())
    output = tf.layers.dense(l, units = 14, name='fc2', kernel_initializer=tf.contrib.layers.xavier_initializer()) # [batch_size, 14]

cross_entropy = tf.nn.sigmoid_cross_entropy_with_logits(labels=labels, logits=output) # cost function
cost = tf.reduce_mean(cross_entropy, name='cost_fn')