Transfer-Learning: 101 Classes

import tensorflow as tf

cifar = tf.keras.datasets.cifar100
(x_train, y_train), (x_test, y_test) = cifar.load_data()
model = tf.keras.applications.ResNet50(
    include_top=True,
    weights=None,
    input_shape=(32, 32, 3),
    classes=100,)

loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)
model.compile(optimizer="adam", loss=loss_fn, metrics=["accuracy"])
model.fit(x_train, y_train, epochs=5, batch_size=64)

import tensorflow as tf
from tensorflow.keras import layers
from tensorflow.keras.models import Sequential

tf.config.get_visible_devices()

import gc
gc.collect()

# !wget https://raw.githubusercontent.com/mrdbourke/tensorflow-deep-learning/main/extras/helper_functions.py

from helper_functions import create_tensorboard_callback, plot_loss_curves, unzip_data, compare_historys, walk_through_dir

# Download data from Google Storage (already preformatted)
# !wget https://storage.googleapis.com/ztm_tf_course/food_vision/101_food_classes_10_percent.zip 

# unzip_data("101_food_classes_10_percent.zip")

# to get a brief "inspection" of the data
# walk_through_dir("101_food_classes_10_percent")

imagesDirPath = "101_food_classes_10_percent/"
trainDirPath = imagesDirPath + 'train/'
testDirPath = imagesDirPath + 'test/'
print(f'TRAIN: {trainDirPath} \n TEST: {testDirPath}')

OUTPUT_IMG_SIZE = (224,224)

trainingData10P = tf.keras.preprocessing.image_dataset_from_directory(trainDirPath,
                                                                                label_mode="categorical",
                                                                                image_size=OUTPUT_IMG_SIZE)
                                                                                
testingData10p = tf.keras.preprocessing.image_dataset_from_directory(testDirPath,
                                                                label_mode="categorical",
                                                                image_size=OUTPUT_IMG_SIZE,
                                                                shuffle=False) # don't shuffle test data for prediction analysis

# Create checkpoint callback to save model for later use
checkpoint_path = "101_classes_10_percent_data_model_checkpoint"
checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(checkpoint_path,
                                                         save_weights_only=True, # save only the model weights
                                                         monitor="val_accuracy", # save the model weights which score the best validation accuracy
                                                         save_best_only=True) # only keep the best model weights on file (delete the rest)

dataAugLayer = Sequential([
  layers.RandomFlip("horizontal"),
  layers.RandomRotation(0.2),
  layers.RandomZoom(0.2),
  layers.RandomHeight(0.2),
  layers.RandomWidth(0.2),
  # preprocessing.Rescaling(1./255) # keep for ResNet50V2, remove for EfficientNetB0
], name ="data_augmentation")

base_model = tf.keras.applications.efficientnet.EfficientNetB0(include_top=False)
base_model.trainable = False

howManyClasses = len(testingData10p.class_names)
howManyClasses

m0Inputs = layers.Input(shape=(224, 224, 3), name="input_layer") # shape of input image
m0Applied = dataAugLayer(m0Inputs) # augment images (only happens during training)
m0Applied = base_model(m0Applied, training=False) # put the base model in inference mode so we can use it to extract features without updating the weights
m0Applied = layers.GlobalAveragePooling2D(name="global_average_pooling")(m0Applied) # pool the outputs of the base model
m0Outputs = layers.Dense(howManyClasses, activation="softmax", name="output_layer")(m0Applied) # same number of outputs as classes
m0 = tf.keras.Model(m0Inputs, m0Outputs)

m0.summary()

m0.compile(loss="categorical_crossentropy",
              optimizer=tf.keras.optimizers.Adam(), # use Adam with default settings
              metrics=["accuracy"])

howManyValidationSteps = int(0.15 * len(testingData10p))
howManyValidationSteps

m0History = m0.fit(trainingData10P,
                                           epochs=5, # fit for 5 epochs to keep experiments quick
                                           validation_data=testingData10p,
                                           validation_steps=int(0.15 * len(testingData10p)), # evaluate on smaller portion of test data
                                           callbacks=[checkpoint_callback])

plot_loss_curves(m0History)

base_model.trainable = True

# RE-Freeze all-but-5 layers
for layer in base_model.layers[:-5]:
  layer.trainable = False

m0.compile(loss='categorical_crossentropy',
              optimizer=tf.keras.optimizers.Adam(1e-4), # 10x lower learning rate than default
              metrics=['accuracy'])

for idx, layer in enumerate(m0.layers):
    if(layer.trainable == True):
        print(f'layer idx {idx}: {layer.name} is trainable')

for idx, layer in enumerate(base_model.layers):
    if(layer.trainable == True):
        print(f'layer idx {idx}: {layer.name} is trainable')

fineTinedEpochCount = 10
m0FineTunedHistory = m0.fit(trainingData10P,
                                                     epochs=fineTinedEpochCount,
                                                     validation_data=testingData10p,
                                                     validation_steps=int(0.15 * len(testingData10p)), # validate on 15% of the test data
                                                     initial_epoch=m0History.epoch[-1]) # start from previous last epoch

compare_historys(original_history=m0History,
                 new_history=m0FineTunedHistory,
                 initial_epochs=5)

m0Loss, m0Accuracy = m0.evaluate(testingData10p)

m0Loss, m0Accuracy

m0.save('food101-10p-transfer-EfficientNetB0')

loadedModel = tf.keras.models.load_model('food101-10p-transfer-EfficientNetB0')

# Check to see if loaded model is a trained model
loaded_loss, loaded_accuracy = loadedModel.evaluate(testingData10p)
loaded_loss, loaded_accuracy

m0PredictionProbabilities = m0.predict(testingData10p, verbose=1)

len(m0PredictionProbabilities)

# should be (numberOfImages, numberOfClasses)
m0PredictionProbabilities.shape

# We get one prediction probability per class
print(f"Number of prediction probabilities for sample 0: {len(m0PredictionProbabilities[0])}")
print(f"What prediction probability sample 0 looks like:\n {m0PredictionProbabilities[0]}")
print(f"The class with the highest predicted probability by the model for sample 0: {m0PredictionProbabilities[0].argmax()}")

# Get JUST THE CLASS NUMBER (0-100) of each label of each prediction
pred_classes = m0PredictionProbabilities.argmax(axis=1)

# How do they look?
pred_classes[:10]

testDataLabels = []
for images, labels in testingData10p.unbatch(): # unbatch the test data and get images and labels
  testDataLabels.append(labels.numpy().argmax()) # append the index which has the largest value (labels are one-hot)
testDataLabels[:10] # check what they look like (unshuffled)

len(testDataLabels)