Keras:ModelCheckpoint和model.fit的verbose有什么差异？
- 1.软件环境⚙️
- 2.问题描述🔍
- 3.解决方法🐡
- 3.1.只打开`model.fit`的`verbose`
- 3.2.只打开`ModelCheckpoint`的`verbose`
- 3.3.同时打开`ModelCheckpoint`和`model.fit`的`verbose`
- 4.结论🤔
⚡插播一条老家自产的糖心苹果，多个品种，欢迎选购！有问题随时私信我⚡：🍎🍎来自雪域高原的馈赠——海拔2000米的大凉山高原生态糖心苹果，欢迎选购！！🍎🍎
大凉山高原生态糖心苹果
1.软件环境⚙️
Windows10教育版64位
Python3.6.3
Tensorflow-GPU2.3.0
Keras2.4.3
CUDA10.1
2.问题描述🔍
我们知道在Keras框架中有多个verbose的设置，比如最常见的model.fit中的verbose：
def fit(self,x=None,y=None,batch_size=None,epochs=1,verbose=1,callbacks=None,validation_split=0.,validation_data=None,shuffle=True,class_weight=None,sample_weight=None,initial_epoch=0,steps_per_epoch=None,validation_steps=None,validation_batch_size=None,validation_freq=1,max_queue_size=10,workers=1,use_multiprocessing=False):"""Trains the model for a fixed number of epochs (iterations on a dataset).Arguments:cverbose: 0, 1, or 2. Verbosity mode.0 = silent, 1 = progress bar, 2 = one line per epoch.Note that the progress bar is not particularly useful whenlogged to a file, so verbose=2 is recommended when not runninginteractively (eg, in a production environment)."""
以及ModelCheckpoint中的verbose：
class EarlyStopping(Callback):def __init__(self,filepath,monitor='val_loss',verbose=0,save_best_only=False,save_weights_only=False,mode='auto',save_freq='epoch',options=None,**kwargs):"""Arguments:verbose: verbosity mode, 0 or 1."""
可以看到verbose主要控制训练模型时，训练的loss、accuracy等评价指标的具体信息：
5/5 [==============================] - ETA: 0s - loss: 3.5273 - accuracy: 0.2000 - precision: 1.0000 - recall: 0.0625
Epoch 00001: val_accuracy improved from -inf to 0.50000, saving model to model_and_log\img_aug+mixup+Xception\Xception-01-3.5273-0.2000-3.3476-0.5000.h5
其实除了上面的verbose，callbacks中还有大量的verbose，比如LearningRateScheduler、EarlyStopping等等都有自己的verbose。
那为什么有多个verbose呢？ 相应的问题就来了：
- 每个
verbose控制的都是同一个东西吗？- 如果同时开启会不会相互覆盖设置？
- 有没有优先级？
3.解决方法🐡
本文以model.fit和ModelCheckpoint的verbose为例，展示了不同情况的verbose会产生什么影响，其它函数的verbose也是相同的结论。
3.1.只打开model.fit的verbose
Epoch 1/30
1/5 [=====>........................] - ETA: 0s - loss: 3.7069 - accuracy: 0.2500 - precision: 0.5000 - recall: 1.0000
2/5 [===========>..................] - ETA: 0s - loss: 3.6293 - accuracy: 0.2500 - precision: 0.7143 - recall: 0.8333
3/5 [=================>............] - ETA: 0s - loss: 3.6301 - accuracy: 0.1667 - precision: 0.6000 - recall: 0.7500
4/5 [=======================>......] - ETA: 0s - loss: 3.6179 - accuracy: 0.1250 - precision: 0.6000 - recall: 0.5000
5/5 [==============================] - ETA: 0s - loss: 3.5854 - accuracy: 0.1000 - precision: 0.6364 - recall: 0.4375
5/5 [==============================] - 5s 900ms/step - loss: 3.5854 - accuracy: 0.1000 - precision: 0.6364 - recall: 0.4375 - val_loss: 3.3600 - val_accuracy: 0.5000 - val_precision: 0.0000e+00 - val_recall: 0.0000e+00
Epoch 2/30
1/5 [=====>........................] - ETA: 0s - loss: 3.4214 - accuracy: 0.2500 - precision: 0.0000e+00 - recall: 0.0000e+00
2/5 [===========>..................] - ETA: 0s - loss: 3.4038 - accuracy: 0.1250 - precision: 0.5000 - recall: 0.1667
3/5 [=================>............] - ETA: 0s - loss: 3.3367 - accuracy: 0.1667 - precision: 0.5000 - recall: 0.1111
4/5 [=======================>......] - ETA: 0s - loss: 3.2700 - accuracy: 0.1875 - precision: 0.5000 - recall: 0.0833
5/5 [==============================] - ETA: 0s - loss: 3.2022 - accuracy: 0.2000 - precision: 0.6667 - recall: 0.1250
5/5 [==============================] - 3s 669ms/step - loss: 3.2022 - accuracy: 0.2000 - precision: 0.6667 - recall: 0.1250 - val_loss: 2.8924 - val_accuracy: 0.5000 - val_precision: 0.0000e+00 - val_recall: 0.0000e+00
Epoch 3/30
1/5 [=====>........................] - ETA: 0s - loss: 2.8710 - accuracy: 0.2500 - precision: 0.0000e+00 - recall: 0.0000e+00
2/5 [===========>..................] - ETA: 0s - loss: 2.8412 - accuracy: 0.1250 - precision: 0.0000e+00 - recall: 0.0000e+00
3/5 [=================>............] - ETA: 0s - loss: 2.8127 - accuracy: 0.1667 - precision: 0.5000 - recall: 0.1000
4/5 [=======================>......] - ETA: 0s - loss: 2.7665 - accuracy: 0.1875 - precision: 0.3333 - recall: 0.1818
5/5 [==============================] - ETA: 0s - loss: 2.7303 - accuracy: 0.1500 - precision: 0.2857 - recall: 0.1429
5/5 [==============================] - 3s 627ms/step - loss: 2.7303 - accuracy: 0.1500 - precision: 0.2857 - recall: 0.1429 - val_loss: 2.3946 - val_accuracy: 0.5000 - val_precision: 0.0000e+00 - val_recall: 0.0000e+00
可以看到，这个时候只会以进度条的方式打印每一个Epoch的评价指标（包括默认的loss、 accuracy、precision、recall以及你自己设定的评价指标），不会打印模型的存储信息。
3.2.只打开ModelCheckpoint的verbose
可以看到，这个时候会根据ModelCheckpoint中设定的监控指标：
checkpoint = ModelCheckpoint(model_x.__name__ + '-{epoch:02d}''-{loss:.4f}''-{accuracy:.4f}''-{val_loss:.4f}''-{val_accuracy:.4f}.h5',monitor='val_accuracy',verbose=1,mode='max',save_best_only=True)
只打印模型的存储信息，模型的评价指标（包括默认的loss、 accuracy、precision、recall以及你自己设定的评价指标）不会打印：
Epoch 00001: val_accuracy improved from -inf to 0.50000, saving model to model_and_log\img_aug+mixup+Xception\Xception-01-3.4481-0.2000-3.2658-0.5000.h5
Epoch 00002: val_accuracy did not improve from 0.50000
Epoch 00003: val_accuracy did not improve from 0.50000
3.3.同时打开ModelCheckpoint和model.fit的verbose
Epoch 1/30
1/5 [=====>........................] - ETA: 0s - loss: 3.5883 - accuracy: 0.5000 - precision: 0.0000e+00 - recall: 0.0000e+00
2/5 [===========>..................] - ETA: 0s - loss: 3.5933 - accuracy: 0.2500 - precision: 1.0000 - recall: 0.1667
3/5 [=================>............] - ETA: 0s - loss: 3.5551 - accuracy: 0.2500 - precision: 0.5000 - recall: 0.1250
4/5 [=======================>......] - ETA: 0s - loss: 3.5576 - accuracy: 0.2500 - precision: 0.6667 - recall: 0.1667
5/5 [==============================] - ETA: 0s - loss: 3.5451 - accuracy: 0.2000 - precision: 0.8000 - recall: 0.2500
Epoch 00001: val_accuracy improved from -inf to 0.50000, saving model to model_and_log\img_aug+mixup+Xception\Xception-01-3.5451-0.2000-3.3282-0.5000.h5
5/5 [==============================] - 4s 889ms/step - loss: 3.5451 - accuracy: 0.2000 - precision: 0.8000 - recall: 0.2500 - val_loss: 3.3282 - val_accuracy: 0.5000 - val_precision: 0.0000e+00 - val_recall: 0.0000e+00
Epoch 2/30
1/5 [=====>........................] - ETA: 0s - loss: 3.3335 - accuracy: 0.2500 - precision: 0.0000e+00 - recall: 0.0000e+00
2/5 [===========>..................] - ETA: 0s - loss: 3.3294 - accuracy: 0.1250 - precision: 0.0000e+00 - recall: 0.0000e+00
3/5 [=================>............] - ETA: 0s - loss: 3.2696 - accuracy: 0.1667 - precision: 0.5000 - recall: 0.1111
4/5 [=======================>......] - ETA: 0s - loss: 3.1904 - accuracy: 0.1875 - precision: 0.6667 - recall: 0.1667
5/5 [==============================] - ETA: 0s - loss: 3.1653 - accuracy: 0.1500 - precision: 0.8000 - recall: 0.2500
Epoch 00002: val_accuracy did not improve from 0.50000
5/5 [==============================] - 3s 599ms/step - loss: 3.1653 - accuracy: 0.1500 - precision: 0.8000 - recall: 0.2500 - val_loss: 2.8628 - val_accuracy: 0.5000 - val_precision: 0.5000 - val_recall: 1.0000
learning_rate= 0.00285
Epoch 3/30
1/5 [=====>........................] - ETA: 0s - loss: 2.8587 - accuracy: 0.2500 - precision: 0.5000 - recall: 0.5000
2/5 [===========>..................] - ETA: 0s - loss: 2.8461 - accuracy: 0.1250 - precision: 0.7500 - recall: 0.5000
3/5 [=================>............] - ETA: 0s - loss: 2.7913 - accuracy: 0.1667 - precision: 0.8000 - recall: 0.4000
4/5 [=======================>......] - ETA: 0s - loss: 2.7635 - accuracy: 0.2500 - precision: 0.6250 - recall: 0.4545
5/5 [==============================] - ETA: 0s - loss: 2.7230 - accuracy: 0.2500 - precision: 0.6364 - recall: 0.5000
Epoch 00003: val_accuracy did not improve from 0.50000
5/5 [==============================] - 3s 617ms/step - loss: 2.7230 - accuracy: 0.2500 - precision: 0.6364 - recall: 0.5000 - val_loss: 2.3710 - val_accuracy: 0.5000 - val_precision: 0.5000 - val_recall: 1.0000
可以看到，如果二者同时打开的话，模型的评价指标和存储信息均会打印，并不会产生影响，二者是完全独立的！
4.结论🤔
model.fit的verbose负责打印每一个Epoch的评价指标（包括默认的loss、accuracy、precision、recall以及你自己设定的评价指标）ModelCheckpoint的verbose负责打印模型的存储信息，如存储路径，存储模型的文件名- 二者是完全独立的不会相互覆盖
渣男！都看到这里了，还不赶紧点赞，评论，收藏走一波？
