python-kerasHow can I use the Python Keras Applications library to build a neural network?

Keras Applications is a library of pre-trained models with deep learning algorithms, developed with a high-level API for building and training models. It can be used to build a neural network in Python.

Example code

from keras.applications import MobileNet

# Create the base model of MobileNet
model = MobileNet(weights='imagenet', include_top=False)

# Add a new top layer
x = model.output
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
x = Dense(1024, activation='relu')(x)
x = Dense(512, activation='relu')(x)
preds = Dense(2, activation='softmax')(x)

# Create a new model
model_2 = Model(inputs=model.input, outputs=preds)

# Freeze the layers
for layer in model_2.layers[:20]:
    layer.trainable=False
for layer in model_2.layers[20:]:
    layer.trainable=True

# Compile the model
model_2.compile(optimizer='Adam', loss='categorical_crossentropy', metrics=['accuracy'])

This code creates a new model based on MobileNet, adds a new top layer, freezes layers, and compiles the model.

Code explanation

1. from keras.applications import MobileNet: imports the MobileNet model from the Keras Applications library.
2. model = MobileNet(weights='imagenet', include_top=False): creates the base model of MobileNet using the pre-trained weights from ImageNet and excluding the top layer.
3. x = model.output: creates a new output layer.
4. x = GlobalAveragePooling2D()(x): applies global average pooling to the output layer.
5. x = Dense(1024, activation='relu')(x): adds a dense layer with 1024 neurons and ReLU activation.
6. x = Dense(1024, activation='relu')(x): adds a dense layer with 1024 neurons and ReLU activation.
7. x = Dense(512, activation='relu')(x): adds a dense layer with 512 neurons and ReLU activation.
8. preds = Dense(2, activation='softmax')(x): adds a dense layer with 2 neurons and softmax activation.
9. model_2 = Model(inputs=model.input, outputs=preds): creates a new model with the inputs from the original model and the new output layer.
10. for layer in model_2.layers[:20]:: freezes the first 20 layers of the model.
11. for layer in model_2.layers[20:]:: allows the remaining layers to be trained.
12. model_2.compile(optimizer='Adam', loss='categorical_crossentropy', metrics=['accuracy']): compiles the model with the Adam optimizer, categorical cross entropy loss function, and accuracy metric.

Helpful links

• Keras Applications Documentation
• MobileNet Documentation
• GlobalAveragePooling2D Documentation
• Dense Layer Documentation

Edit this code on GitHub