import math from pathlib import Path 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, GlobalAveragePooling2D) from tensorflow.keras.optimizers.legacy import Adam, Adadelta from tensorflow.keras.losses import categorical_crossentropy sys.path.append('./resources/libraries') import ei_tensorflow.training WEIGHTS_PATH = './transfer-learning-weights/keras/mobilenet_v2_weights_tf_dim_ordering_tf_kernels_0.35_96.h5' # Download the model weights root_url = 'https://cdn.edgeimpulse.com/' p = Path(WEIGHTS_PATH) if not p.exists(): print(f"Pretrained weights {WEIGHTS_PATH} unavailable; downloading...") if not p.parent.exists(): p.parent.mkdir(parents=True) weights_data = requests.get(root_url + WEIGHTS_PATH[2:]).content with open(WEIGHTS_PATH, 'wb') as f: f.write(weights_data) print(f"Pretrained weights {WEIGHTS_PATH} unavailable; downloading OK") print("") INPUT_SHAPE = (96, 96, 3) base_model = tf.keras.applications.MobileNetV2( input_shape = INPUT_SHAPE, alpha=0.35, weights = WEIGHTS_PATH ) 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(Reshape((-1, model.layers[-1].output.shape[3]))) model.add(Dense(8, activation='relu')) model.add(Dropout(0.1)) model.add(Flatten()) model.add(Dense(classes, activation='softmax')) # Implements the data augmentation policy def augment_image(image, label): # Flips the image randomly image = tf.image.random_flip_left_right(image) # Increase the image size, then randomly crop it down to # the original dimensions resize_factor = random.uniform(1, 1.2) new_height = math.floor(resize_factor * INPUT_SHAPE[0]) new_width = math.floor(resize_factor * INPUT_SHAPE[1]) image = tf.image.resize_with_crop_or_pad(image, new_height, new_width) image = tf.image.random_crop(image, size=INPUT_SHAPE) # Vary the brightness of the image image = tf.image.random_brightness(image, max_delta=0.2) return image, label train_dataset = train_dataset.map(augment_image, num_parallel_calls=tf.data.AUTOTUNE) BATCH_SIZE = 32 EPOCHS = args.epochs or 60 LEARNING_RATE = args.learning_rate or 0.01 train_dataset = train_dataset.batch(BATCH_SIZE, drop_remainder=False) validation_dataset = validation_dataset.batch(BATCH_SIZE, drop_remainder=False) callbacks.append(BatchLoggerCallback(BATCH_SIZE, train_sample_count, epochs=EPOCHS)) model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=LEARNING_RATE), loss='categorical_crossentropy', metrics=['accuracy']) model.fit(train_dataset, validation_data=validation_dataset, epochs=EPOCHS, verbose=2, callbacks=callbacks, class_weight=ei_tensorflow.training.get_class_weights(Y_train)) 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, class_weight=ei_tensorflow.training.get_class_weights(Y_train) )