#!/usr/bin/env python3 """ Transfer learning... """ import tensorflow.keras as K def preprocess_data(X, Y): """ Data preprocessing """ # Preprocesses a tensor or Numpy array encoding a batch of images. X_p = K.applications.resnet50.preprocess_input(x=X) # Converts a class vector (integers) to binary class matrix Y_p = K.utils.to_categorical(y=Y) return X_p, Y_p if __name__ == '__main__': """ making transfer learning """ # Loads the CIFAR10 dataset (X_train, Y_train), (X_test, Y_test) = K.datasets.cifar10.load_data() # Load weights pre-trained on ImageNet. # Do not include the ImageNet classifier at the top. base_model = K.applications.InceptionResNetV2(weights="imagenet", input_shape=(299, 299, 3), include_top=False) base_model.summary() # Preprocessing data X_train, Y_train = preprocess_data(X_train, Y_train) X_test, Y_test = preprocess_data(X_test, Y_test) # Freeze the base_model # base_model.trainable = False # Create new model on top inputs = K.Input(shape=(32, 32, 3)) # Resizes the images contained in a 4D tensor. input_image = K.layers.Lambda( lambda image: K.backend.resize_images( x=image, height_factor=299/32, width_factor=299/32, data_format='channels_last' ) )(inputs) # Add to base model layer = base_model(input_image, training=False) # Global average pooling operation for spatial data. poolG1 = K.layers.GlobalAveragePooling2D()(layer) # stack new 3 layer network fc1 = K.layers.Dense(units=512)(poolG1) batchN1 = K.layers.BatchNormalization()(fc1) activation1 = K.layers.Activation('relu')(batchN1) dropout1 = K.layers.Dropout(0.3)(activation1) fc2 = K.layers.Dense(units=512)(dropout1) batchN2 = K.layers.BatchNormalization()(fc2) activation2 = K.layers.Activation('relu')(batchN2) dropout2 = K.layers.Dropout(0.3)(activation2) fc3 = K.layers.Dense(10, activation='softmax')(dropout2) model = K.Model(inputs=inputs, outputs=fc3) # Freeze the layers model base_model.trainable = False for layer in base_model.layers: layer.trainable = False # Compile the model should be done *after* setting layers to non-trainable # optimizer optimizer = K.optimizers.Adam() model.compile(optimizer=optimizer, loss="categorical_crossentropy", metrics=['accuracy']) model.fit(x=X_train, y=Y_train, batch_size=300, shuffle=True, epochs=2, validation_data=(X_test, Y_test), verbose=True) # Unfreeze selected layers from IRNV2 for layer in base_model.layers[:498]: layer.trainable = False for layer in base_model.layers[498:]: layer.trainable = True print("\nTraining bottom couple hundred layers\n") # optimizer with low learning rate optimizer = K.optimizers.Adam(1e-5) # Training the top layer model.compile(optimizer=optimizer, loss="categorical_crossentropy", metrics=['accuracy']) history = model.fit(x=X_train, y=Y_train, validation_data=(X_test, Y_test), batch_size=300, epochs=5, verbose=True) model.save('cifar10.h5') print("saved to ./cifar10.h5 NEVER STOP SMILING! ")

#!/usr/bin/env python3 import tensorflow.keras as K # preprocess_data = __import__('0-transfer').preprocess_data # to fix issue with saving keras applications K.learning_phase = K.backend.learning_phase _, (X, Y) = K.datasets.cifar10.load_data() X_p, Y_p = preprocess_data(X, Y) model = K.models.load_model('cifar10.h5') model.evaluate(X_p, Y_p, batch_size=128, verbose=1)