Keras,附加回调日志

Question

我有一个回调计算on_epoch_end验证数据中的几个额外指标和测试数据的每10个时期.

我还有一个CSVLogger将正常指标保存到日志文件的回调.

从我的回调中有一种简单的方法可以将一两列添加到正确编写的日志中CSVLogger吗？

Answer 1

您可以将其他指标插入字典中logs.

from keras.callbacks import Callback

class ComputeMetrics(Callback):
    def on_epoch_end(self, epoch, logs):
        logs['val_metric'] = epoch ** 2  # replace it with your metrics
        if (epoch + 1) % 10 == 0:
            logs['test_metric'] = epoch ** 3  # same
        else:
            logs['test_metric'] = np.nan

请记住CSVLogger在fit通话之前放置此回叫.稍后出现在列表中的回调将收到修改后的版本logs.例如,

model = Sequential([Dense(1, input_shape=(10,))])
model.compile(loss='mse', optimizer='adam')
model.fit(np.random.rand(100, 10),
          np.random.rand(100),
          epochs=30,
          validation_data=(np.random.rand(100, 10), np.random.rand(100)),
          callbacks=[ComputeMetrics(), CSVLogger('1.log')])

现在,如果你看一看输出日志文件,你会看到两个附加列test_metric和val_metric:

epoch,loss,test_metric,val_loss,val_metric
0,0.547923130989,nan,0.370979120433,0
1,0.525437340736,nan,0.35585285902,1
2,0.501358469725,nan,0.341958616376,4
3,0.479624577463,nan,0.329370084703,9
4,0.460121934414,nan,0.317930338383,16
5,0.440655426979,nan,0.307486981452,25
6,0.422990380526,nan,0.298160370588,36
7,0.406809270382,nan,0.289906248748,49
8,0.3912438941,nan,0.282540213466,64
9,0.377326357365,729,0.276457450986,81
10,0.364721306562,nan,0.271435074806,100
11,0.353612961769,nan,0.266939682364,121
12,0.343238875866,nan,0.263228923082,144
13,0.333940329552,nan,0.260326927304,169
14,0.325931007862,nan,0.25773427248,196
15,0.317790198028,nan,0.255648627281,225
16,0.310636150837,nan,0.25411529541,256
17,0.304091459513,nan,0.252928718328,289
18,0.298703012466,nan,0.252127869725,324
19,0.292693507671,6859,0.251701972485,361
20,0.287824733257,nan,0.251610517502,400
21,0.283586999774,nan,0.251790778637,441
22,0.27927801609,nan,0.252100949883,484
23,0.276239238977,nan,0.252632959485,529
24,0.273072380424,nan,0.253150621653,576
25,0.270296501517,nan,0.253555388451,625
26,0.268056542277,nan,0.254015884399,676
27,0.266158599854,nan,0.254496408701,729
28,0.264166412354,nan,0.254723013639,784
29,0.262506003976,24389,0.255338237286,841