Keras training aids and optimization tools

Original: http://blog.csdn.net/zzulp/article/details/76591341 1 Callbacks

Callbacks provides a series of classes that can be called back during training to enable observation and interference in the training process. In addition to some classes provided by the library, users can customize the class. The following lists the more useful callback classes.

modelcheckpoint

class name	effect	constructor
is used to save the model	modelcheckpoint (filepath, monitor= ' Val_loss ') between epochs, save_best_only =false, Save_weights_only=false, mode= ' auto ', period=1)
earlystopping	when early Stop is activated (if a loss is found to be less than the previous epoch training), the training is stopped after the patience epoch.	earlystopping (monitor= ' Val_loss ', patience=0, mode= ' auto ')
tensorboard	generate TB-required logs	Tensorboard (log_dir= './logs ', histogram_freq=0, Write_graph=true, Write_images=false, Embeddings_freq=0, Embeddings_layer_names=none, Embeddings_metadata=none)
Reducelronplateau	when the indicator becomes Reduce learning rate	Reducelronplateau (monitor= ' Val_loss ', factor=0.1, patience=10, mode= ' auto ', epsilon=0.0001, CoolD Own=0, min_lr=0)

Example:

From keras.callbacks import modelcheckpoint

model = sequential ()
model.add (Dense, input_dim=784, kernel_ initializer= ' uniform '))
Model.add (Activation (' Softmax '))
model.compile (loss= ' categorical_crossentropy ') , optimizer= ' Rmsprop ')

checkpointer = Modelcheckpoint (filepath= "/tmp/weights.h5", Save_best_only=true)
TENSBRD = Tensorboard (logdir= ' Path/of/log ')
model.fit (X_train, Y_train, batch_size=128, Callbacks=[checkpointer , TENSBRD])

1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10

PS: After adding the Tensorboard callback class, you can use the TensorFlow tensorboard command line to open the Visual Web service. 2 Application

This module provides a pre-trained image model based on image-net to facilitate our migration learning. When first used, the model weight data is downloaded to the ~/.keras/models directory.

Image Model	Description	constructor Function
InceptionV3		InceptionV3 (include_top=true, weights= ' imagenet ', input_tensor=none,input_shape=none,pooling=none,classes=1000)
ResNet50		ResNet50 (include_top=true, weights= ' imagenet ', input_tensor=none,input_shape=none,pooling=none,classes=1000)
VGG19		VGG19 (include_top=true, weights= ' imagenet ', input_tensor=none,input_shape=none,pooling=none,classes=1000)
VGG16		VGG16 (include_top=true, weights= ' imagenet ', input_tensor=none,input_shape=none,pooling=none,classes=1000)
Xception		Xception (include_top=true, weights= ' imagenet ', Input_tensor=none,input_shape=none,pooling=none, classes=1000)

Parameter description

Parameters	Description
Include_top	Whether to keep the top-level fully connected network, false as long as the bottleneck
Weights	' Imagenet ' stands for load pre-training weights, none for random initialization
Input_tensor	Can be filled in keras tensor as the model of the image output tensor
Input_shape	A tuple with a length of 3, indicating the shape of the input image, the width of the picture must be greater than 197
Pooling	Feature extraction network pooling mode. None represents non-pooling, and the output of the last convolutional layer is 4D tensor. ' AVG ' represents the global average pooling, ' Max ' represents the global maximum value pooling
Classes	Number of categories in the picture category, available when Include_top=true Weight=none

For migration learning, you can refer to this article: How to implement image classification on very small datasets. It describes the process of transforming images and using existing models and fine-tune new classifiers. 3 Visualization of models

The Plot_model function is provided in the Utils package, which is used to present a model in the form of an image. This feature relies on Pydot-ng and Graphviz.
Pip Install Pydot-ng Graphviz

From keras.utils import Plot_model
model = keras.applications.InceptionV3 ()
Plot_model (model, to_file= ' Model.png ')

1 2 3 1 2 3