import tensorflow as tf from tensorflow.keras import Model from tensorflow.keras.models import Sequential from tensorflow.keras.layers import Dense, InputLayer, Dropout, Conv1D, Flatten, Reshape, MaxPooling1D, BatchNormalization, Conv2D, GlobalMaxPooling2D, Lambda from tensorflow.keras.optimizers import Adam, Adadelta from tensorflow.keras.losses import categorical_crossentropy # Load best model from initial training sys.path.append('./resources/libraries') import ei_tensorflow.training INPUT_SHAPE = (96, 96, 3) base_model = tf.keras.applications.MobileNetV2( input_shape=INPUT_SHAPE, alpha=0.35, weights='./transfer-learning-weights/keras/mobilenet_v2_weights_tf_dim_ordering_tf_kernels_0.35_96.h5', include_top=True ) base_model.trainable = False model = Sequential() model.add(InputLayer(input_shape=INPUT_SHAPE, name='x_input')) # Don't include the base model's top layers last_layer_index = -3 model.add(Model(inputs=base_model.inputs, outputs=base_model.layers[last_layer_index].output)) model.add(Dense(16)) model.add(Dropout(0.1)) model.add(Flatten()) model.add(Dense(classes, activation='softmax')) model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0005), loss='categorical_crossentropy', metrics=['accuracy']) BATCH_SIZE = 32 train_dataset, validation_dataset = ei_tensorflow.training.set_batch_size(BATCH_SIZE, train_dataset, validation_dataset) callbacks.append(BatchLoggerCallback(BATCH_SIZE, train_sample_count)) model.fit(train_dataset, validation_data=validation_dataset, epochs=10, verbose=2, callbacks=callbacks) print('') print('Initial training done.', flush=True) # How many epochs we will fine tune the model FINE_TUNE_EPOCHS = 10 # What percentage of the base model's layers we will fine tune FINE_TUNE_PERCENTAGE = 65 print('Fine-tuning best model for {} epochs...'.format(FINE_TUNE_EPOCHS), flush=True) # Load best model from initial training model = ei_tensorflow.training.load_best_model(BEST_MODEL_PATH) # Determine which layer to begin fine tuning at model_layer_count = len(model.layers) fine_tune_from = math.ceil(model_layer_count * ((100 - FINE_TUNE_PERCENTAGE) / 100)) # Allow the entire base model to be trained model.trainable = True # Freeze all the layers before the 'fine_tune_from' layer for layer in model.layers[:fine_tune_from]: layer.trainable = False model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.000045), loss='categorical_crossentropy', metrics=['accuracy']) model.fit(train_dataset, epochs=FINE_TUNE_EPOCHS, verbose=2, validation_data=validation_dataset, callbacks=callbacks)