我的分类器损失很大,准确率始终为 0

My classifier has very big loss and accuracy is always 0

我正在训练分类器以获得因子进行优化。 我的数据集一开始包含 800 个样本(有些相似,只是稍作修改)。

我使用 GoogleColab 环境使用 TensorFlow 开发了我的模型。

我使用了一个简单的 MLP 来解决这个问题,3 个隐藏层 每个层有 256 个节点作为第一阶段。我也有 64 类 .

我有 可变长度 输入,我已经用 "-1" 填充.

解决了这个问题

以我的实际特征我知道我会得到不好的准确率,但我没想到零准确率非常大的损失 .

这是我的数据集,省略了一些我注意到会对准确性产生负面影响的特征:

     0  1   2   4  5  6  8  9   11  13    15      17    19      21
805  6  10  11  1  3  1  6  64  2   1.0   64.0    64.0  64.0    -1.0
334  6  12  18  0  2  4  7  2   1   32.0  128.0   64.0  128.0   -1.0
781  7  10  11  1  3  1  6  2   2   2.0   64.0    32.0  32.0    64.0

[已编辑]:这是我的一些标签行:

    0
0   108
1   30
2   30
3   16
4   62
5   126
6   22
7   30
8   48

这是我的一组结果:

epoch[0] step [0] train -- loss : 50751.734375, accuracy : 0.0
epoch[0] step [100] train -- loss : 27310.064453125, accuracy : 0.0
epoch[0] step [200] train -- loss : 58120.6015625, accuracy : 0.0
epoch[0] step [300] train -- loss : 31801.9453125, accuracy : 0.0
epoch[0] step [400] train -- loss : 54360.76171875, accuracy : 0.0
epoch[0] step [500] train -- loss : 59946.67578125, accuracy : 0.0
epoch[1] step [0] train -- loss : 40612.06640625, accuracy : 0.0
epoch[1] step [100] train -- loss : 43229.734375, accuracy : 0.0
epoch[1] step [200] train -- loss : 36951.84375, accuracy : 0.0
epoch[1] step [300] train -- loss : 45225.828125, accuracy : 0.0
epoch[1] step [400] train -- loss : 47055.1796875, accuracy : 0.0
epoch[1] step [500] train -- loss : 54023.23046875, accuracy : 0.0

我想知道为什么我会得到这么大的损失,而且我的损失 也没有收敛 :(.

那些是我得到的准确率和损失的奇怪图表:

我的代码运行得很好,但我仍然怀疑是不是我写得不好所以弄得一团糟。

这是我代码的重要部分:

class MLP():
    '''
      This is the implementation of the Multi Layer Perceptron
    '''
    def __init__(self, x_train, y_train, n_classes, n_hiddens=3, activation=tf.nn.relu):
      ...       
    ################################ Create the model ##############################    
    def multilayer_perceptron(self,X):      
        # Hidden fully connected layer with n_hidden_1 neurons
        layer_1 = tf.layers.dense(inputs=X, units= self.n_hidden_1, use_bias=True, kernel_initializer=self._init, name= 'layer_1')
        layer_1 = tf.layers.batch_normalization(layer_1,training=self.is_train)
        layer_1 = self.activation(layer_1)     

        # Hidden fully connected layer with n_hidden_2 neurons
        layer_2 = tf.layers.dense(inputs=layer_1, units= self.n_hidden_2, use_bias=True, kernel_initializer=self._init, name= 'layer_2')
        layer_2 = tf.layers.batch_normalization(layer_2,training=self.is_train)
        layer_2 = self.activation(layer_2)   

        # Hidden fully connected layer with n_hidden_3 neurons
        layer_3 = tf.layers.dense(inputs=layer_2, units= self.n_hidden_3, use_bias=True, kernel_initializer=self._init, name= 'layer_3')
        layer_3 = tf.layers.batch_normalization(layer_3, training=self.is_train)
        layer_3 = self.activation(layer_3)  

        # Output fully connected layer with the output
        out_layer = tf.layers.dense(inputs=layer_3, units= self.n_classes, use_bias=True, kernel_initializer=self._init, name= 'out_layer')

        tf.summary.histogram('pre-activations', out_layer) 
        return layer_1, layer_2, layer_3, out_layer  

在这里我计算了训练集的损失和准确性:

 def loss(self, X, Y):
        _, _, _, self.predicted_out = self.multilayer_perceptron(X) 
        print("Predicted out", self.predicted_out)
        with tf.name_scope('loss'):          
                loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=self.predicted_out, labels=Y))
        tf.summary.scalar('loss', loss) 

        with tf.name_scope('accuracy'):
                predicted_class = tf.nn.softmax(self.predicted_out)
                with tf.name_scope('correct_prediction'):
                    correct_prediction = tf.equal(tf.argmax(predicted_class, 1), tf.argmax(Y, 1))
                with tf.name_scope('accuracy'):
                    accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float'))
        tf.summary.scalar('accuracy', accuracy)
        self.merged =  tf.summary.merge_all()

        return loss, accuracy 

这是训练的功能:

def train(self):
        self.train_writer = tf.summary.FileWriter('./Graph', self.sess.graph) 
    # training data
        train_input = self.Normalize(self.x_train)
        train_output = self.y_train.copy()

        save_sess=self.sess

      #costs history :
        costs = []
        costs_inter=[] 
      #for early stopping :
        best_cost=1000000
        stop = False
        last_improvement=0

        n_samples = train_input.shape[0] # size of the training set

       #train the mini_batches model using the early stopping criteria
        epoch = 0
        while epoch < self.max_epochs and stop == False:
            #train the model on the traning set by mini batches
            #suffle then split the training set to mini-batches of size self.batch_size
            seq =list(range(n_samples))
            random.shuffle(seq)
            mini_batches = [
                seq[k:k+self.batch_size]
                for k in range(0,n_samples, self.batch_size)
            ]

            avg_cost = 0. # The average cost of mini_batches
            step= 0

            for sample in mini_batches:                  
                batch_x = x_train.iloc[sample, :]
                batch_y =train_output.iloc[sample, :]
                feed_dict={self.X: batch_x,self.Y:batch_y, self.is_train:True}
                self.train_summary, _, cost,acc=self.sess.run([self.merged, self.train_step, self.loss_, self.accuracy_], feed_dict=feed_dict)
                avg_cost += cost *len(sample)/n_samples      
                print('epoch[{}] step [{}] train -- loss : {}, accuracy : {}'.format(epoch,step, cost, acc))
                self.train_writer.add_summary(self.train_summary, global_step=step)
                step += 100

            #cost history since the last best cost
            costs_inter.append(avg_cost)
            epoch +=1

        #Test the model 
        pred = tf.nn.softmax(self.predicted_out)  # Apply softmax to logits
        correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(self.Y, 1))
        # Calculate accuracy
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
        accuracy_test = self.sess.run([accuracy],feed_dict={self.X: x_test, self.Y: y_test,self.is_train:False})
        #print("Accuracy:", accuracy.eval({self.X: x_test, self.Y: y_test}))
        print("Accuracy_test : ", accuracy_test)

        # Writes the summaries to disk
        self.train_writer.flush()
        # Flushes the summaries to disk and closes the SummaryWriter
        self.train_writer.close()                              
        return costs

在这里我调用我的方法 train,我打算应用交叉验证来查看我是否可以提高模型的准确性,但我还没有:

def cross_validation(self,batch_size, n_hidden_1 , n_hidden_2, n_hidden_3, learning_rate):
        ##### Other parameter 
        self.batch_size = batch_size
        self.n_hidden_1 = n_hidden_1
        self.n_hidden_2 = n_hidden_2
        self.n_hidden_3 = n_hidden_3
        self.learning_rate = learning_rate
        self.require_improvement= 20
        self.max_epochs = 80
        self._init = tf.random_normal_initializer
        self.optimizer=tf.train.AdamOptimizer      
        loss = 0   
        tf.reset_default_graph()

        with tf.name_scope('input'):
          self.X=tf.placeholder("float",shape=[None,self.x_train.shape[1]])
          self.Y=tf.placeholder("float",shape=[None,self.y_train.shape[1]])
          self.is_train = tf.placeholder(tf.bool, name="is_train") 
        self.loss_, self.accuracy_ = self.loss(self.X, self.Y)
        self.train_step = self.optimizer(self.learning_rate).minimize(self.loss_)

        # Initiate a tensor session
        init = tf.global_variables_initializer()
        self.sess = tf.Session()
        self.sess.run(init)

        #train the model 
        loss = self.train()

        self.sess.close()
        del self.sess
        return  loss  

最后这是我的主要内容:

  if __name__=='__main__':
  tbc = tb.TensorBoardColab()
  mlp =   MLP(x_train, y_train, n_classes)
  loss= mlp.cross_validation(batch_size, n_hidden_1 , n_hidden_2, n_hidden_3, learning_rate) 

我原以为精度会很差,但不是“0”,而且没有任何收敛的巨大损失让我震惊。谁能告诉我这里的问题在哪里?

有几点需要注意

  1. 你应该在每次训练开始前删除tf摘要文件,因为根据你的代码全局步骤将从0重新开始

  2. 你的损失函数是 softmax_cross_entropy_with_logits_v2,要使用这个你可能需要在 onehot 中编码你的标签,并尝试使用内部 softmax 函数最小化靠近那个 onehot 标签的 logit 层功能。如果您想保留当前的真实标签,请勾选 sparse_softmax_cross_entropy_with_logits. The usages are similar but some of them need to be onehot label. Check detailed explaination