使用Keras编写复杂的loss时一般会将loss作为模型的一个层，则模型的输入包括原始输入和y_true，模型输出即为loss，例如yolov3keras版本的代码就是将loss作为模型的一层，因其计算loss比较复杂。下面介绍一个简单的例子，利用神经网络预测Iris数据集
需要的包
import keras.backend as K
from keras import layers, Model
import numpy as np
from keras.utils import plot_model, to_categorical
首先加载并划分数据集，Iris共有4个特征，最后一列为label（减一是为了从0开始），文末附上数据集
data = np.loadtxt('Iris.txt')
np.random.shuffle(data)
x_train, y_train = data[:100, :4], data[:100, 4]-1 # (100,4) (100,)
x_valid, y_valid = data[100:, :4], data[100:, 4]-1 # (50,4) (50,)
将label转换为onehot
y_train = to_categorical(y_train, num_classes=3) # (100,3)
y_valid = to_categorical(y_valid, num_classes=3) # (50,3)
构建预测网络，网络输出3为的向量，即3个类别的概率
def net():
inputs = layers.Input(shape=[4, ], name='feature_input')
dense1 = layers.Dense(32, activation='relu', name='dense1')(inputs)
out = layers.Dense(3, activation='linear', name='out1')(dense1)
return Model(inputs, out)
接下来编写自己的loss函数，传入的args为[y_true, y_pred]，这里实现了一个简单的交叉熵。
def my_loss(args, batch_size, C):
"""
args: [y_true, y_pred], (batch_size, C)
batch_size:
C: classes num.
"""
y_true, y_pred = args[0], args[1]
log_pred = K.log(y_pred)
loss = -K.sum(y_true*log_pred, axis=1)
batch_tensor = K.cast(batch_size, dtype=float)
return loss/batch_tensor
构建训练模型，loss_body即为loss层，arguments是预先传入需要的参数，train_model为训练模型，输入包括两个，一个是原始输入一个是y_true，输出为loss，注意compile时lambda y_true, y_pred:的顺序别反了，否则模型模型跟踪不到梯度，会报错，fit的时候y要喂入一个和x大小一样的0数组或其它的也行，反正它不起作用，但一定要喂入。保存模型时我们保留预测模型而不是训练模型，并且使用预测模型来预测。
# build model
predict_body = net()
y_true = layers.Input(shape=[3, ]) # one hot label
loss_body = layers.Lambda(my_loss, output_shape=(1,), name='my_loss',
arguments={'batch_size': 16, 'C': 3})([y_true, predict_body.output])
train_model = Model(inputs=[predict_body.input, y_true], outputs=loss_body)
train_model.compile(optimizer='adam',
loss={'my_loss': lambda y_true, y_pred: y_pred}) # y_true, y_pred别反了, 因为模型送入的输入就是这样,否则跟踪不了梯度
plot_model(train_model, show_shapes=True)
train_model.fit(x=[x_train, y_train], y=np.zeros(len(x_train)), batch_size=16, epochs=20,
validation_data=([x_valid, y_valid], np.zeros(len(x_valid))))
predict_body.save_weights('predict.h5') # 保存预测模型
用plot_model画出train_model如下：
模型开始训练：
不过这样有个缺点，在训练过程中我们无法看到准确率，只能看到loss，因此可以做如下修改：
将模型变为多输出，输出层除了loss（out2）外再加上预测模型的输出（out1），预测模型的输出用来计算准确率不参与loss计算，因为keras会将两个loss相加，所以out1的loss直接返回0。metric计算out1的accuracy即可。由于算accuracy又要用到y_true，所以在fit的时候y传入[np.zeros(len(x_valid)), y_valid]，前面的zeros用来传给loss，y_valid用来传给metric计算准确率。
import keras.backend as K
from keras import layers, Model
import numpy as np
from keras.utils import plot_model, to_categorical
def net():
inputs = layers.Input(shape=[4, ], name='feature_input')
dense1 = layers.Dense(32, activation='relu', name='dense1')(inputs)
out = layers.Dense(3, activation='linear', name='out1')(dense1)
return Model(inputs, out)
def my_loss(args, batch_size, C):
"""
args: [y_true, y_pred], (batch_size, C)
batch_size:
C: classes num.
"""
y_true, y_pred = args[0], args[1]
soft_pred = K.softmax(y_pred)
log_pred = K.log(soft_pred+0.0000001)
loss = -K.sum(y_true*log_pred, axis=1)
batch_tensor = K.cast(batch_size, dtype=float)
return loss/batch_tensor
data = np.loadtxt('Iris.txt')
np.random.shuffle(data)
x_train, y_train = data[:100, :4], data[:100, 4]-1 # (100,4) (100,)
x_valid, y_valid = data[100:, :4], data[100:, 4]-1 # (50,4) (50,)
y_train = to_categorical(y_train, num_classes=3)
y_valid = to_categorical(y_valid, num_classes=3)
# build model
predict_body = net()
y_true = layers.Input(shape=[3, ], name='y_true') # one hot label
loss_body = layers.Lambda(my_loss, output_shape=(1,), name='out2',
arguments={'batch_size': 16, 'C': 3})([y_true, predict_body.output])
train_model = Model(inputs=[predict_body.input, y_true],
outputs=[loss_body, predict_body.output])
train_model.compile(optimizer='adam',
loss={'out2': lambda y_true, y_pred: y_pred,
'out1': lambda y_true, y_pred: K.constant(0)},
metrics={'out1': 'accuracy'})
train_model.fit(x=[x_train, y_train],
y=[np.zeros(len(x_train)), y_train], batch_size=16, epochs=50,
validation_data=[
[x_valid, y_valid],
[np.zeros(len(x_valid)), y_valid]])
训练过程就可以看到accuracy了
虽然这种写法不太优雅，但还是能看到metric了
Iris.txt
5.1 3.5 1.4 0.2 1
4.9 3.0 1.4 0.2 1
4.7 3.2 1.3 0.2 1
4.6 3.1 1.5 0.2 1
5.0 3.6 1.4 0.2 1
5.4 3.9 1.7 0.4 1
4.6 3.4 1.4 0.3 1
5.0 3.4 1.5 0.2 1
4.4 2.9 1.4 0.2 1
4.9 3.1 1.5 0.1 1
5.4 3.7 1.5 0.2 1
4.8 3.4 1.6 0.2 1
4.8 3.0 1.4 0.1 1
4.3 3.0 1.1 0.1 1
5.8 4.0 1.2 0.2 1
5.7 4.4 1.5 0.4 1
5.4 3.9 1.3 0.4 1
5.1 3.5 1.4 0.3 1
5.7 3.8 1.7 0.3 1
5.1 3.8 1.5 0.3 1
5.4 3.4 1.7 0.2 1
5.1 3.7 1.5 0.4 1
4.6 3.6 1.0 0.2 1
5.1 3.3 1.7 0.5 1
4.8 3.4 1.9 0.2 1
5.0 3.0 1.6 0.2 1
5.0 3.4 1.6 0.4 1
5.2 3.5 1.5 0.2 1
5.2 3.4 1.4 0.2 1
4.7 3.2 1.6 0.2 1
4.8 3.1 1.6 0.2 1
5.4 3.4 1.5 0.4 1
5.2 4.1 1.5 0.1 1
5.5 4.2 1.4 0.2 1
4.9 3.1 1.5 0.1 1
5.0 3.2 1.2 0.2 1
5.5 3.5 1.3 0.2 1
4.9 3.1 1.5 0.1 1
4.4 3.0 1.3 0.2 1
5.1 3.4 1.5 0.2 1
5.0 3.5 1.3 0.3 1
4.5 2.3 1.3 0.3 1
4.4 3.2 1.3 0.2 1
5.0 3.5 1.6 0.6 1
5.1 3.8 1.9 0.4 1
4.8 3.0 1.4 0.3 1
5.1 3.8 1.6 0.2 1
4.6 3.2 1.4 0.2 1
5.3 3.7 1.5 0.2 1
5.0 3.3 1.4 0.2 1
7.0 3.2 4.7 1.4 2
6.4 3.2 4.5 1.5 2
6.9 3.1 4.9 1.5 2
5.5 2.3 4.0 1.3 2
6.5 2.8 4.6 1.5 2
5.7 2.8 4.5 1.3 2
6.3 3.3 4.7 1.6 2
4.9 2.4 3.3 1.0 2
6.6 2.9 4.6 1.3 2
5.2 2.7 3.9 1.4 2
5.0 2.0 3.5 1.0 2
5.9 3.0 4.2 1.5 2
6.0 2.2 4.0 1.0 2
6.1 2.9 4.7 1.4 2
5.6 2.9 3.6 1.3 2
6.7 3.1 4.4 1.4 2
5.6 3.0 4.5 1.5 2
5.8 2.7 4.1 1.0 2
6.2 2.2 4.5 1.5 2
5.6 2.5 3.9 1.1 2
5.9 3.2 4.8 1.8 2
6.1 2.8 4.0 1.3 2
6.3 2.5 4.9 1.5 2
6.1 2.8 4.7 1.2 2
6.4 2.9 4.3 1.3 2
6.6 3.0 4.4 1.4 2
6.8 2.8 4.8 1.4 2
6.7 3.0 5.0 1.7 2
6.0 2.9 4.5 1.5 2
5.7 2.6 3.5 1.0 2
5.5 2.4 3.8 1.1 2
5.5 2.4 3.7 1.0 2
5.8 2.7 3.9 1.2 2
6.0 2.7 5.1 1.6 2
5.4 3.0 4.5 1.5 2
6.0 3.4 4.5 1.6 2
6.7 3.1 4.7 1.5 2
6.3 2.3 4.4 1.3 2
5.6 3.0 4.1 1.3 2
5.5 2.5 4.0 1.3 2
5.5 2.6 4.4 1.2 2
6.1 3.0 4.6 1.4 2
5.8 2.6 4.0 1.2 2
5.0 2.3 3.3 1.0 2
5.6 2.7 4.2 1.3 2
5.7 3.0 4.2 1.2 2
5.7 2.9 4.2 1.3 2
6.2 2.9 4.3 1.3 2
5.1 2.5 3.0 1.1 2
5.7 2.8 4.1 1.3 2
6.3 3.3 6.0 2.5 3
5.8 2.7 5.1 1.9 3
7.1 3.0 5.9 2.1 3
6.3 2.9 5.6 1.8 3
6.5 3.0 5.8 2.2 3
7.6 3.0 6.6 2.1 3
4.9 2.5 4.5 1.7 3
7.3 2.9 6.3 1.8 3
6.7 2.5 5.8 1.8 3
7.2 3.6 6.1 2.5 3
6.5 3.2 5.1 2.0 3
6.4 2.7 5.3 1.9 3
6.8 3.0 5.5 2.1 3
5.7 2.5 5.0 2.0 3
5.8 2.8 5.1 2.4 3
6.4 3.2 5.3 2.3 3
6.5 3.0 5.5 1.8 3
7.7 3.8 6.7 2.2 3
7.7 2.6 6.9 2.3 3
6.0 2.2 5.0 1.5 3
6.9 3.2 5.7 2.3 3
5.6 2.8 4.9 2.0 3
7.7 2.8 6.7 2.0 3
6.3 2.7 4.9 1.8 3
6.7 3.3 5.7 2.1 3
7.2 3.2 6.0 1.8 3
6.2 2.8 4.8 1.8 3
6.1 3.0 4.9 1.8 3
6.4 2.8 5.6 2.1 3
7.2 3.0 5.8 1.6 3
7.4 2.8 6.1 1.9 3
7.9 3.8 6.4 2.0 3
6.4 2.8 5.6 2.2 3
6.3 2.8 5.1 1.5 3
6.1 2.6 5.6 1.4 3
7.7 3.0 6.1 2.3 3
6.3 3.4 5.6 2.4 3
6.4 3.1 5.5 1.8 3
6.0 3.0 4.8 1.8 3
6.9 3.1 5.4 2.1 3
6.7 3.1 5.6 2.4 3
6.9 3.1 5.1 2.3 3
5.8 2.7 5.1 1.9 3
6.8 3.2 5.9 2.3 3
6.7 3.3 5.7 2.5 3
6.7 3.0 5.2 2.3 3
6.3 2.5 5.0 1.9 3
6.5 3.0 5.2 2.0 3
6.2 3.4 5.4 2.3 3
5.9 3.0 5.1 1.8 3
