首先辨析一下概念:
1. loss是整体网络进行优化的目标, 是需要参与到优化运算,更新权值W的过程的
2. metric只是作为评价网络表现的一种“指标”, 比如accuracy,是为了直观地了解算法的效果,充当view的作用,并不参与到优化过程
一、keras自定义损失函数
在keras中实现自定义loss, 可以有两种方式,一种自定义 loss function, 例如:
# 方式一
def vae_loss(x, x_decoded_mean):
xent_loss = objectives.binary_crossentropy(x, x_decoded_mean)
kl_loss = - 0.5 * K.mean(1 + z_log_sigma - K.square(z_mean) - K.exp(z_log_sigma), axis=-1)
return xent_loss + kl_loss
vae.compile(optimizer='rmsprop', loss=vae_loss)
或者通过自定义一个keras的层(layer)来达到目的, 作为model的最后一层,最后令model.compile中的loss=None:
# 方式二
# Custom loss layer
class CustomVariationalLayer(Layer):
def __init__(self, **kwargs):
self.is_placeholder = True
super(CustomVariationalLayer, self).__init__(**kwargs)
def vae_loss(self, x, x_decoded_mean_squash):
x = K.flatten(x)
x_decoded_mean_squash = K.flatten(x_decoded_mean_squash)
xent_loss = img_rows * img_cols * metrics.binary_crossentropy(x, x_decoded_mean_squash)
kl_loss = - 0.5 * K.mean(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)
return K.mean(xent_loss + kl_loss)
def call(self, inputs):
x = inputs[0]
x_decoded_mean_squash = inputs[1]
loss = self.vae_loss(x, x_decoded_mean_squash)
self.add_loss(loss, inputs=inputs)
# We don't use this output.
return x
y = CustomVariationalLayer()([x, x_decoded_mean_squash])
vae = Model(x, y)
vae.compile(optimizer='rmsprop', loss=None)
在keras中自定义metric非常简单,需要用y_pred和y_true作为自定义metric函数的输入参数 点击查看metric的设置
注意事项:
1. keras中定义loss,返回的是batch_size长度的tensor, 而不是像tensorflow中那样是一个scalar
2. 为了能够将自定义的loss保存到model, 以及可以之后能够顺利load model, 需要把自定义的loss拷贝到keras.losses.py 源代码文件下,否则运行时找不到相关信息,keras会报错
有时需要不同的sample的loss施加不同的权重,这时需要用到sample_weight,例如
discriminator.train_on_batch(imgs, [valid, labels], class_weight=class_weights)
相关参考:https://www.jianshu.com/p/209e43b62d79
二、keras中的样本权重
# Import
import numpy as np
from sklearn.utils import class_weight
# Example model
model = Sequential()
model.add(Dense(32, activation='relu', input_dim=100))
model.add(Dense(1, activation='sigmoid'))
# Use binary crossentropy loss
model.compile(optimizer='rmsprop',
loss='binary_crossentropy',
metrics=['accuracy'])
# Calculate the weights for each class so that we can balance the data
weights = class_weight.compute_class_weight('balanced',
np.unique(y_train),
y_train)
# Add the class weights to the training
model.fit(x_train, y_train, epochs=10, batch_size=32, class_weight=weights)
Note that the output of the class_weight.compute_class_weight()
is an numpy array like this: [2.57569845 0.68250928]
.