tensowflow keras - model.predict 给出所有相同的输出

Question

1. 型。预测给出所有相同的输出

2. 验证和测试精度< 60%

我尝试将最后一个隐藏层更改为softmax，但它仍然没有解决这个问题。任何反馈将不胜感激。我也尝试过使用超参数，但我仍然找不到任何修复程序。

raw_cvs_data = np.loadtxt('data_to_train.csv',delimiter=',')

 raw_cvs_data_to_compute = np.loadtxt('data_to_compute.csv',delimiter=',')

 unscaled_inputs_all = raw_cvs_data[:,1:]

 targets_all = raw_cvs_data[:,0]

 inputs_to_compute = raw_cvs_data_to_compute[:]

 predicted_target=[]

 # balancing the dataset

 num_one_targets = int(np.sum(targets_all)) # count how many targets are 1

 zero_targets_counter = 0 # counter for target 0

 indices_to_remove = [] # remove extra input/target pairs for balance 

 # count the number of targets 0, when get same amount of target 1 and 0, make entries where target is zero

 for i in range(targets_all.shape[0]):

         if targets_all[i] == 0:

             zero_targets_counter +=1

             if zero_targets_counter > num_one_targets:

                 indices_to_remove.append(i)

 unscaled_inputs_equal_priors = np.delete(unscaled_inputs_all,indices_to_remove, axis = 0)

 targets_equal_priors = np.delete(targets_all, indices_to_remove, axis = 0)

 #Shuffle the data 

 shuffled_indices = np.arange(scaled_inputs.shape[0])

 np.random.shuffle(shuffled_indices) #shuffle pairs

 shuffled_inputs = scaled_inputs[shuffled_indices]

 shuffled_targets = targets_equal_priors[shuffled_indices]

 # splitting data

 samples_count = shuffled_inputs.shape[0]

 # |training|validation|testing| 80-10-10

 train_samples_count = int(0.8 * samples_count)

 validation_samples_count = int(0.1 *samples_count)

 test_samples_count = samples_count - train_samples_count - validation_samples_count

 train_inputs = shuffled_inputs[:train_samples_count]

 train_targets = shuffled_targets[:train_samples_count]

Answer 1

您的问题来自以下代码行：。tf.keras.layers.Dense(output_size, activation='sigmoid')问题是你正在使用2个神经元的“”激活，而不是1个神经元。sigmoid使用 或 .2 neurons + activation = 'softmax'1 neuron + activation='sigmoid'

Answer 2

如果，看起来，您正在尝试使用单热编码标签进行分类，那么您的代码存在两个问题。首先，你使用了错误的损失;均方误差 （MSE） 用于回归问题，而不是分类问题。更改模型编译以使用二进制交叉熵损失，即：model.compile(optimizer='sgd',&nbsp;loss='binary_crossentropy',&nbsp;metrics=['accuracy'])其次，正如其他答案中已经提示的那样，将最后一层的激活函数更改为 ，即：softmaxtf.keras.layers.Dense(output_size,&nbsp;activation='softmax')这是单热编码标签的正确方法。