TypeError：在Tensorflow中使用自定义度量时，类型为“Tensor”的对象没有len（）

Question

我正在开发一个多类分类问题（4类）的模型使用Keras和Tensorflow后端。y_test的值具有二维格式：

0 1 0 0
0 0 1 0
0 0 1 0

def my_metric(targ, predict):
    val_predict = predict
    val_targ = tf.math.argmax(targ, axis=1)
    return metrics.balanced_accuracy_score(val_targ, val_predict)

hidden_neurons = 50
timestamps = 20
nb_features = 18

model = Sequential()

model.add(LSTM(
                units=hidden_neurons,
                return_sequences=True, 
                input_shape=(timestamps,nb_features),
                dropout=0.15
                #recurrent_dropout=0.2
              )
         )

model.add(TimeDistributed(Dense(units=round(timestamps/2),activation='sigmoid')))

model.add(Dense(units=hidden_neurons,
               activation='sigmoid'))


model.add(Flatten())

model.add(Dense(units=nb_classes,
               activation='softmax'))

model.compile(loss="catehtml" target="_blank">gorical_crossentropy",
              metrics = [my_metric],
              optimizer='adadelta')

---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

~/Anaconda3/lib/python3.6/site-packages/keras/engine/training.py in compile（self,optimizer,loss,metrics,loss_weights,sample_weight_mode,weighted_metrics,target_tensors,**kwargs)449 output_metrics=nested_metrics[i]450 output_weighted_metrics=nested_weighted_metrics[i]-->451 handle_metrics（output_metrics）452 handle_metrics（output_weighted_metrics,weights=weights）453

~/Anaconda3/lib/python3.6/site-packages/keras/engine/training.py in handle_metrics（metrics,weights）418 metric_result=weighted_metric_fn(y_true,y_pred,419 weights=weights,-->420 mask=masks[i])421 422#追加到self.metrics_names,self.metric_tensors,

~/Anaconda3/lib/python3.6/site-packages/keras/engine/training_utils.py in weighted(y_true,y_pred,weights,mask)402“”“403#score_array具有ndim>=2-->404 score_array=fn(y_true,y_pred)405如果mask不是none：406#将掩码转换为floatX以避免float64在Theano中向上转换

在my_metric（targ,predict）22 val_predict=predict 23 val_targ=tf.math.argmax（targ,axis=1)--->24返回metric.balanced_accuracy_score（val_targ,val_predict）25#return 5 26

~/Anaconda3/lib/python3.6/site-packages/sklearn/metrics/classification.py in balanced_accuracy_score(y_true,y_pred,sample_weight,adjusted）
1431 1432“”->1433 C=confusion_matrix(y_true,y_pred,sample_weight=sample_weight)1434 with np.errstate(Divide='ignore',invalid='ignore'):1435
per_class=np.diag(C)/C.sum（axis=1)

~/Anaconda3/lib/python3.6/site-packages/sklearn/metrics/classification.py in confusion_matrix(y_true,y_pred,labels,sample_weight)251 252“”-->253 y_type,y_true,y_pred=_check_targets(y_true,y_pred)254如果y_type不在（“binary”,“multiclass”）中：255提高值错误（“%s不受支持”%y_type)

~/Anaconda3/lib/python3.6/site-packages/sklearn/metrics/classification.py in_check_targets(y_true,y_pred)69 y_pred:数组或指示器矩阵70“”--->71 check_consistent_length(y_true,y_pred)72 type_true=type_of_target(y_true)73 type_pred=type_of_target(y_pred)

~/Anaconda3/lib/python3.6/site-packages/sklearn/utils/validation.py in check_consistent_length(*arrays)229“”“230-->231 lengths=[_num_samples(X)for X in arrays（如果X不是None）]232 uniques=np.unique（长度）233如果len(uniques)>1:

~/Anaconda3/lib/python3.6/site-packages/sklearn/utils/validation.py in(.0)229“”“230-->231 lengths=[_num_samples(X)for X in arrays（如果X不是None）]232 uniques=np.unique（长度）233如果len(uniques)>1:

~/Anaconda3/lib/python3.6/site-packages/sklearn/utils/validation.py in_num_samples(x)146返回x.shape[0]147 else：-->148返回len(x)149 else：150返回len(x)

TypeError：类型为“Tensor”的对象没有len（）

Answer 1

不能在Keras张量上调用sklearn函数。您需要自己使用Keras的后端函数来实现该功能，如果您使用的是TF后端，则需要使用TensorFlow函数来实现该功能。

balanced_accuracy_score定义为在每列中获得的查全率的平均值。检查此链接以了解精度和召回的实现情况。对于balanced_accuracy_score，您可以按照以下方式实现它：

import keras.backend as K

def balanced_recall(y_true, y_pred):
    """
    Computes the average per-column recall metric
    for a multi-class classification problem
    """ 
    true_positives = K.sum(K.round(K.clip(y_true * y_pred, 0, 1)), axis=0)  
    possible_positives = K.sum(K.round(K.clip(y_true, 0, 1)), axis=0)   
    recall = true_positives / (possible_positives + K.epsilon())    
    balanced_recall = K.mean(recall)
    return balanced_recall