Tensorflow

Transfer learning with VGG and Keras for image classification task # Credits: # Author: Gabriel Cassimiro # Blog post: https://towardsdatascience.com/transfer-learning-with-vgg16-and-keras-50ea161580b4 # GitHub Repo: https://github.com/gabrielcassimiro17/object-detection # import tensorflow_datasets as tfds from tensorflow.keras import layers, models from tensorflow.keras.utils import to_categorical from tensorflow.keras.callbacks import EarlyStopping ## Loading images and labels (train_ds, train_labels), (test_ds, test_labels) = tfds.load( "tf_flowers", split=["train[:70%]", "train[:30%]"], ## Train test split batch_size=-1, as_supervised=True, # Include labels ) ## Resizing images train_ds = tf.image.resize(train_ds, (150, 150)) test_ds = tf.image.resize(test_ds, (150, 150)) ## Transforming labels to correct format train_labels = to_categorical(train_labels, num_classes=5) test_labels = to_categorical(test_labels, num_classes=5) from tensorflow.keras.applications.vgg16 import VGG16 from tensorflow.keras.applications.vgg16 import preprocess_input ## Loading VGG16 model base_model = VGG16(weights="imagenet", include_top=False, input_shape=train_ds[0].shape) base_model.trainable = False ## Not trainable weights ## Preprocessing input train_ds = preprocess_input(train_ds) test_ds = preprocess_input(test_ds) flatten_layer = layers.Flatten() dense_layer_1 = layers.Dense(50, activation='relu') dense_layer_2 = layers.Dense(20, activation='relu') prediction_layer = layers.Dense(5, activation='softmax') model = models.Sequential([ base_model, flatten_layer, dense_layer_1, dense_layer_2, prediction_layer ]) model.compile( optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'], ) es = EarlyStopping(monitor='val_accuracy', mode='max', patience=5, restore_best_weights=True) model.fit(train_ds, train_labels, epochs=50, validation_split=0.2, batch_size=32, callbacks=[es]) model.evaluate(test_ds, test_labels)

Useful resources for learning and creating a Tensorflow Dataset. See also code snippets below. Building a data pipeline: https://cs230.stanford.edu/blog/datapipeline/ tf.data API, Build TensorFlow input pipelines: https://www.tensorflow.org/guide/data tf.data API, Consuming sets of files: https://www.tensorflow.org/guide/data#consuming_sets_of_files Better performance with the tf.data API: https://www.tensorflow.org/guide/data_performance Keras Sequence Generator: https://stanford.edu/~shervine/blog/keras-how-to-generate-data-on-the-fly Loading and Preprocessing Data with TensorFlow: https://canvas.education.lu.se/courses/3766/pages/chapter-13-loading-and-preprocessing-data-with-tensorflow?module_item_id=109789 tf.data.Dataset generators with parallelization: the easy way: https://medium.com/@acordier/tf-data-dataset-generators-with-parallelization-the-easy-way-b5c5f7d2a18 Feed numpy files (.npz) which contain features X and a label y: import tensorflow as tf import numpy as np def read_npy_file(item): x, y = np.load(item.decode()) return x.astype(np.float32), file_list = ['/foo/bar.npz', '/foo/baz.npz'] dataset = tf.data.Dataset.from_tensor_slices(file_list) dataset = dataset.map( lambda item: tuple( tf.py_func(func=read_npy_file, inp=[item], Tout=[tf.float32,]) ) ) # Read numpy files (.npz), extract labels and return a new tf.data.Dataset def get_dataset(file_names_list, num_classes=2): """Creates a new TensorFlow Dataset ---------- Parameters: file_names_list: list of file paths num_classes: int Returns: (Tensor, Tensor) """ # Load the numpy files def map_func(file_path): np_data = np.load(file_path) x_data = np_data["x"] y_label = np_data["y"] return x_data.astype(np.float32), tf.one_hot(indices=y_label, depth=num_classes) # Map function numpy_func = lambda item: tf.numpy_function(map_func, [item], [tf.float32, tf.float32]) # Create a new tensorflow dataset dataset = tf.data.Dataset.from_tensor_slices(file_list) # Use map to load the numpy files in parallel dataset = dataset.map(numpy_func, num_parallel_calls=tf.data.AUTOTUNE) return dataset The following code snippet has been taken from https://medium.com/@acordier/tf-data-dataset-generators-with-parallelization-the-easy-way-b5c5f7d2a18 ...