ronzillia

# balance data via loss for each batch
x_train, y_train,index = batch_data
label=np.argmax(y_train,axis=1)
num_pos=np.count_nonzero(label)
num_neg=len(label)-num_pos
pos_weight=np.true_divide(num_neg+1,len(label)+1)
neg_weight=np.true_divide(num_pos+1,len(label)+1)
class_weight=np.array([[neg_weight,pos_weight]])
# plus one to avoid all-zero and all-one problem
_ ,train_cost= sess.run([train_op,classification_loss_op], feed_dict={input_x: x_train,input_y_classification: y_train,bs_holder:batch_size,training_flag:True,tf_class_weight:class_weight})

ronzillia

def classification_loss(self,logit,input_y_classification,class_weight):
        labels=input_y_classification
        
        weight_per_label = tf.transpose( tf.matmul(labels, tf.transpose(class_weight)) ) #shape [1, batch_size]
                                                   # this is the weight for each datapoint, depending on its label
        entropy=tf.nn.softmax_cross_entropy_with_logits(logits=logit, labels=labels, name="xent_raw")                                           
        xent = tf.multiply(weight_per_label, entropy) #shape [1, batch_size]
        cost = tf.reduce_mean(xent) #shape 1
        
        self._summaries['classification_loss'] = tf.summary.scalar('classification_loss', cost)

ronzillia

# after the whole epoch
 
tp=conf_mtx[1][1]
fp=conf_mtx[0][1]
tn=conf_mtx[0][0]
fn=conf_mtx[1][0]
precision=np.truediv(tp,tp+fp)
recall=np.true_divide(tp,tp +fn)
tn_rate=np.true_divide(tn,tn+fp)

ronzillia

con_mtx = tf.confusion_matrix(labels, prediction,num_classes=self.num_class)