What is the use of verbose in Keras while validating the model?

问题

I'm running the LSTM model for the first time. Here is my model:

opt = Adam(0.002)
inp = Input(...)
print(inp)
x = Embedding(....)(inp)
x = LSTM(...)(x)
x = BatchNormalization()(x)
pred = Dense(5,activation='softmax')(x)

model = Model(inp,pred)
model.compile(....)

idx = np.random.permutation(X_train.shape[0])
model.fit(X_train[idx], y_train[idx], nb_epoch=1, batch_size=128, verbose=1)

What is the use of verbose while training the model?

回答1:

Check documentation for model.fit here.

By setting verbose 0, 1 or 2 you just say how do you want to 'see' the training progress for each epoch.

verbose=0 will show you nothing (silent)

verbose=1 will show you an animated progress bar like this:

verbose=2 will just mention the number of epoch like this:

回答2:

verbose: Integer. 0, 1, or 2. Verbosity mode.

Verbose=0 (silent)

Verbose=1 (progress bar)

Train on 186219 samples, validate on 20691 samples
Epoch 1/2
186219/186219 [==============================] - 85s 455us/step - loss: 0.5815 - acc: 
0.7728 - val_loss: 0.4917 - val_acc: 0.8029
Train on 186219 samples, validate on 20691 samples
Epoch 2/2
186219/186219 [==============================] - 84s 451us/step - loss: 0.4921 - acc: 
0.8071 - val_loss: 0.4617 - val_acc: 0.8168

Verbose=2 (one line per epoch)

Train on 186219 samples, validate on 20691 samples
Epoch 1/1
 - 88s - loss: 0.5746 - acc: 0.7753 - val_loss: 0.4816 - val_acc: 0.8075
Train on 186219 samples, validate on 20691 samples
Epoch 1/1
 - 88s - loss: 0.4880 - acc: 0.8076 - val_loss: 0.5199 - val_acc: 0.8046

回答3:

For verbose > 0, fit method logs:

loss: value of loss function for your training data
acc: accuracy value for your training data.

Note: If regularization mechanisms are used, they are turned on to avoid overfitting.

if validation_data or validation_split arguments are not empty, fit method logs:

val_loss: value of loss function for your validation data
val_acc: accuracy value for your validation data

Note: Regularization mechanisms are turned off at testing time because we are using all the capabilities of the network.

For example, using verbose while training the model helps to detect overfitting which occurs if your acc keeps improving while your val_acc gets worse.