这对我的训练集有什么负面影响？
我试图训练我的模型，我的成本输出减少了每个时期，直到它达到接近于零的值，然后变成负值我想知道具有负成本输出的含义是什么？
Cost after epoch 0: 3499.608553
Cost after epoch 1: 2859.823284
Cost after epoch 2: 1912.205967
Cost after epoch 3: 1041.337282
Cost after epoch 4: 385.100483
Cost after epoch 5: 19.694999
Cost after epoch 6: 0.293331
Cost after epoch 7: 0.244265
Cost after epoch 8: 0.198684
Cost after epoch 9: 0.156083
Cost after epoch 10: 0.117224
Cost after epoch 11: 0.080965
Cost after epoch 12: 0.047376
Cost after epoch 13: 0.016184
Cost after epoch 14: -0.012692
Cost after epoch 15: -0.039486
Cost after epoch 16: -0.064414
Cost after epoch 17: -0.087688
Cost after epoch 18: -0.109426
Cost after epoch 19: -0.129873
Cost after epoch 20: -0.149069
Cost after epoch 21: -0.169113
Cost after epoch 22: -0.184217
Cost after epoch 23: -0.200351
Cost after epoch 24: -0.215847
Cost after epoch 25: -0.230574
Cost after epoch 26: -0.245604
Cost after epoch 27: -0.259469
Cost after epoch 28: -0.272469
Cost after epoch 29: -0.284447
我正在使用张量流进行训练，它是一个带有2个隐藏层的简单神经网络，learning_rate = 0.0001，number_of_epoch = 30，mini-batch_size = 50，train-test-ratio = 69/29并且所有数据集都是101434训练样例Cost使用交叉熵方程进行计算
tf.nn.sigmoid_cross_entropy_with_logits(logits=Z3, labels=Y)
这意味着标签不符合成本函数期望它们的格式。
传递给sigmoid_cross_entropy_with_logits每个标签应该为0或1（对于二元分类）或包含0和1（对于2个以上类）的向量。 否则，它不会按预期工作。
对于n类，输出层应该有n单位，并且标签应该在传递给sigmoid_cross_entropy_with_logits之前进行编码：
Y = tf.one_hot(Y, n)
这假设Y是从0到n-1的标签列表或一维标签数组。
上一篇: what does it mean having a negative cost for my training set?