def level2():
# Destroying the current TF graph - https://keras.io/backend/
backend.clear_session()
print("\n### LEVEL2 ###")
conv_base2 = NASNetLarge(weights='imagenet', include_top=False, input_shape=(img_height, img_width, 3))
model2 = models.Sequential()
model2.add(conv_base2)
model2.add(layers.GlobalAveragePooling2D())
model2.add(layers.Dense(fcLayer1, activation='relu'))
model2.add(layers.Dropout(dropout))
model2.add(layers.Dense(classes, activation='softmax'))
print("\nLoading trained weights from " + weights_path + " ...")
model2.load_weights(weights_path)
# unfreezing the base network up to a specific layer:
if freezeUptoLayer == "":
conv_base2.trainable = True
print ("\ntrainable layers: ",int(len(model2.trainable_weights) / 2))
else:
print("\ntrainable layers before unfreezing the base network up to " + freezeUptoLayer + ": ",int(len(model2.trainable_weights) / 2)) # weights = weights + bias = 2 pro layer
conv_base2.trainable = True
set_trainable = False
for layer in conv_base2.layers:
if layer.name == freezeUptoLayer: set_trainable = True
if set_trainable: layer.trainable = True
else: layer.trainable = False
print("trainable layers after the base network unfreezed from layer " + freezeUptoLayer + ": ", int(len(model2.trainable_weights)/2))
print("\nLEVEL2 Model after unfreezing the base network")
model2.summary()
model2.compile(loss='categorical_crossentropy', optimizer=optimizers.RMSprop(lr=learning_rate/10, decay=lr_decay), metrics=['acc'])
print ("\n### Validating ... ")
val_loss, val_acc = model2.evaluate_generator(test_generator, steps=nbrTestImages, verbose=0)
print('Validation Results before training unfreeze layers and trained densely connected layers:\nValidation loss:',val_loss,",",'Validation accuracy:', val_acc, "\n")
# Jointly training both the unfreeze layers and the added trained densely connected layers
callbacks_list_L2 = [ModelCheckpoint(filepath=model_path+'l2.h5', save_weights_only=False, monitor='val_acc', verbose=1, save_best_only=True),
ReduceLROnPlateau(monitor='val_acc', factor=factorL2, patience=patiencel2, verbose=1),
TensorBoard(log_dir=TensorBoardLogDir+'\\level2')]
print ("\n### Level2 Training ... ")
history = model2.fit_generator(
train_generator,
steps_per_epoch=(nbrTrainImages * classes) // (batch * 5),
epochs=epochsL2,
callbacks=callbacks_list_L2,
validation_data=test_generator,
validation_steps=nbrTestImages,
verbose=verbose_train)
history_val2 = [history.history] # saving all results of the final test
plot(history_val2, "LEVEL2:", epochsL2)
print("\n###LEVEL2 Training finished successfully ###")
def level1():
# Building the model using the pretrained model
conv_base1 = NASNetLarge(weights='imagenet', include_top=False, input_shape=(img_height, img_width, 3))
print("\n### LEVEL1 ###\npretrained network:")
conv_base1.summary()
model = models.Sequential()
model.add(conv_base1)
model.add(layers.GlobalAveragePooling2D())
model.add(layers.Dense(fcLayer1, activation='relu'))
model.add(layers.Dropout(dropout))
model.add(layers.Dense(classes, activation='softmax'))
# freezing the base network
print("trainable layers bevor freezing:", int(len(model.trainable_weights)/2)) # weights = weights + bias = 2 pro layer
conv_base1.trainable = False
print("trainable layers after freezing:", int(len(model.trainable_weights)/2))
print("\npretrained network + densely connected classifier")
model.summary()
# training the added layers only
model.compile(loss='categorical_crossentropy', optimizer=optimizers.RMSprop(lr=learning_rate, decay=lr_decay), metrics=['acc'])
callbacks_list_L1 = [ModelCheckpoint(filepath=weights_path, save_weights_only=True, monitor='val_acc', verbose=1, save_best_only=True),
ReduceLROnPlateau(monitor='val_acc', factor=factorL1, patience=patiencel1, verbose=1),
TensorBoard(log_dir=TensorBoardLogDir+'\\level1')]
print("\n### Level1 Training ... ")
# training the model
history = model.fit_generator(
train_generator,
steps_per_epoch=(nbrTrainImages * classes) // (batch * 5),
epochs=epochsL1,
callbacks=callbacks_list_L1,
validation_data=test_generator,
validation_steps=nbrTestImages,
verbose=verbose_train)
history_val1 = [history.history] # saving all results of the final test
plot(history_val1, "LEVEL1:", epochsL1)
print("\n### LEVEL1 Training finished successfully ###")
print("\nLoading trained weights from " + weights_path + " ...")
model.load_weights(weights_path)
model.compile(loss='categorical_crossentropy', optimizer=optimizers.RMSprop(lr=learning_rate), metrics=['acc'])
print("\n### Saving Level1 Model to ", model_path+'l1.h5', " ... ")
model.save(model_path+'l1.h5')
"""
#Import NASNetLarge architecture with imagenet weights
from keras.applications import NASNetLarge
conv_base = NASNetLarge(weights='imagenet', include_top=False, input_shape=(350,350,3))
from keras import models, layers
model = models.Sequential()
model.add(conv_base)
model.add(layers.Flatten())
model.add(layers.Dense(256, activation = 'relu'))
model.add(layers.Dense(3, activation = 'softmax'))
#v0, convolutional layers 17-18 free
conv_base.trainable = True
set_trainable = False
for layer in conv_base.layers:
if layer.name == 'normal_conv_1_18' or layer.name == 'normal_conv_1_17':
set_trainable = True
if set_trainable:
layer.trainable = True
else:
layer.trainable = False
#Optimizer = Adam
model.compile(loss='categorical_crossentropy', optimizer=optimizers.Adam(lr=1e-5), metrics=['acc'])
#folder with training dataset
train_dir = '/tolkach/training_dataset_SN/'