def run_experiment(n, large_data_set=False, generator=False):
"""
Run an experiment. One experiment loads the dataset, trains the model, and outputs the evaluation metrics after
training.
:param n: Number of experiments to perform
:param large_data_set: Set to True of you want to save the large dataset to hard disk and use generator for training
:param generator: Set to True is you want to use a generator to train the network.
:return: n/a
"""
for experiments in range(n):
if large_data_set:
save_data()
else:
data_train, data_test, data_val, info = load_data()
data_train, data_test, data_val,\
train_images, train_labels, test_images, test_labels,\
val_images, val_labels = preprocess_data(data_train, data_test, data_val)
visualize(data_train, data_test, info)
# Fit the model on the training data and validate on the validation data.
if generator:
train_images = np.load(
'/home/mike/Documents/Alexnet_Client_Backend/file_names.npy'
)
train_labels = np.load(
'/home/mike/Documents/Alexnet_Client_Backend/oh_labels.npy'
)
train_data = DataGenerator(train_images, train_labels,
batch_size)
else:
# Make the images, label paris into the tf.data.Dataset, shuffle the data and specify its batch size.
train_data = tf.data.Dataset.from_tensor_slices(
(train_images, train_labels))
train_data = train_data.repeat().shuffle(6149).batch(100)
test_data = tf.data.Dataset.from_tensor_slices(
(test_images, test_labels))
test_data = test_data.repeat().shuffle(1020).batch(batch_size)
val_data = tf.data.Dataset.from_tensor_slices(
(val_images, val_labels))
val_data = val_data.batch(batch_size)
# With the three parameters, the client script calls the AlexNet model(as a class) in alexnet.py
model = AlexNet(train_data, test_data, val_data)
# Compile the model using Adam optimizer and categorical_crossentropy loss function.
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['acc', top_5_acc])
# Give the model structure summary after complete the above-mentioned calling.
model.summary()
if generator:
file_names = np.load(data_path + 'file_names.npy')
num_files = file_names.shape[0]
del file_names
model.fit_generator(generator=train_data,
steps_per_epoch=int(num_files //
batch_size),
epochs=epochs,
verbose=verbose,
validation_data=val_data)
else:
model.fit(train_data,
epochs=epochs,
validation_data=val_data,
verbose=verbose,
steps_per_epoch=steps_per_epoch)
predictions(model, test_images, test_labels, num_examples=5)
# Evaluate the model
loss, accuracy, top_5 = model.evaluate(test_data,
verbose=verbose,
steps=5)
# Append the metrics to the scores lists in case you are performing an experiment which involves comparing
# training over many iterations.
loss_scores.append(loss)
acc_scores.append(accuracy)
top_5_acc_scores.append(top_5)
# Print the mean, std, min, and max of the validation accuracy scores from your experiment.
print(acc_scores)
print('Mean_accuracy={}'.format(np.mean(acc_scores)),
'STD_accuracy={}'.format(np.std(acc_scores)))
print('Min_accuracy={}'.format(np.min(acc_scores)),
'Max_accuracy={}'.format(np.max(acc_scores)))
# Get the batch shape
for data_batch, label_batch in train_generator:
print("data batch shape:", data_batch.shape)
print("label batch shape:", label_batch)
break
# Set the Tensorbaord
log_dir = "logs/fit/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir)
callback_list = [tensorboard_callback]
# Train the model
history = model.fit(train_generator,
steps_per_epoch=train_num // BATCH_SIZE,
epochs=EPOCHS,
validation_data=validation_generator,
validation_steps=val_num // BATCH_SIZE)
# Save the model
# -model.save('/home/mike/Documents/image_gesture/leapGestRecog_small_categorical.h5')
# Evaluate the model with visulizing the result
acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']
epochs = range(len(acc))
plt.figure(figsize=(15, 4))
plt.subplot(1, 2, 1)
# The terminal shows the number of images belonging to 2 classes.
train_num = train_generator.samples
valid_num = valid_generator.samples
# Need to start the following command in Ubuntu Terminal after executing the script.
# tensorboard --logdir logs/fit
log_dir = "logs/fit/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir)
callback_list = [tensorboard_callback]
# Set up the verbose=1 (or verbose=1) for visible (or invisible) epochs.
model.fit(train_generator,
epochs=EPOCHS,
steps_per_epoch=train_num // BATCH_SIZE,
validation_data=valid_generator,
validation_steps=valid_num // BATCH_SIZE,
callbacks=callback_list,
verbose=1)
# The system saves the whole model into the direcotry: /home/mike/Documents/AlexNet-tf2/content. The
# model of my_model.h5 has the quite big size of 748.6 MB.
model.save(model_dir)
# To view the diagrams, users need to upload the Python script into Jupyter Notebook and run the
# the script or directly upload and run the original ipython script.
class_names = ['bike', 'ship']
x_valid, label_batch = next(iter(valid_generator))
prediction_values = model.predict_classes(x_valid)
# The plot will be realized in the Jupyter Notebook after running the script in either Python or
base_path = "dataset"
train_h5_file = "food_c101_n10099_r64x64x3.h5"
test_h5_file = "food_test_c101_n1000_r64x64x3.h5"
X_train, y_train, X_test, y_test = read_from_h5(base_path, train_h5_file,
test_h5_file)
from keras.callbacks import ModelCheckpoint
checkpoint = ModelCheckpoint('best_model.h5',
monitor='val_loss',
save_best_only=True,
mode='auto')
steps_per_epoch = int(len(y_train) / BATCH_SIZE) # 300
validation_steps = int(len(y_test) / BATCH_SIZE) # 90
# Training
model_info = model.fit(X_train,
y_train,
validation_data=(X_test, y_test),
batch_size=BATCH_SIZE,
shuffle="batch",
epochs=NUM_EPOCH,
verbose=1)
model.save("food_classification.h5")
# plot model history after each epoch
from visulization import plot_model_history
plot_model_history(model_info)
import os.path
import wget
from sklearn.model_selection import train_test_split
from dataset import Dataset
from alexnet import AlexNet
print('Loading data..')
images, labels = Dataset('res/data1', AlexNet.SCALE_SIZE, mean_image=AlexNet.MEAN_IMAGE).load()
print('Splitting into train&validation..')
X_train, X_val, y_train, y_val = train_test_split(images, labels)
print('Looking for pre-trained weights..')
weights_file = 'res/alexnet-caffemodel.npy'
if not os.path.exists(weights_file):
wget.download('https://www.dropbox.com/s/ekgz9jtj1ybtxmj/alexnet-caffemodel.npy?dl=1', 'res')
print('Fine-tuning AlexNet')
alex_net = AlexNet(labels.shape[1], weights_file)
alex_net.fit(X_train, X_val, y_train, y_val, freeze=True, epochs=1000, lr=0.001)
class_mode='categorical')
train_num = train_generator.samples
# Please start the following tensorboard in the Ubuntu Terminal after executing the script.
# $ Tensorboard --logdir logs/fit
log_dir="logs/fit/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=log_dir)
callback_list = [tensorboard_callback]
# Set verbose=1 (or verbose=0) for visibale (or invisible) training procedure.
model.fit(train_generator,
epochs=EPOCHS,
steps_per_epoch=train_num//BATCH_SIZE,
callbacks=callback_list,
verbose=1)
# It is the test generator as similar as the above.
test_datagen = ImageDataGenerator(rescale=1.0/255)
test_generator = test_datagen.flow_from_directory(test_dir,
target_size=(image_width,image_height),
class_mode='categorical')
test_num = test_generator.samples
# Evalute the trained model and return both loss and the test accuracy.
EPOCHS = 5
MODEL_NAME = 'simple-alexnet-epoch{}.model'.format(EPOCHS)
print("[INFO] loading DATA...")
dataset = np.load(argv[1])
labelset = np.load(argv[2])
WIDTH = dataset.shape[1]
HEIGHT = dataset.shape[2]
data = dataset[:, :, :, np.newaxis]
(trainData, testData, trainLabels,
testLabels) = train_test_split(data, labelset, test_size=0.1)
trainLabels = np_utils.to_categorical(trainLabels)
testLabels = np_utils.to_categorical(testLabels)
print("[INFO] compiling model...")
model = AlexNet(width=WIDTH, height=HEIGHT, lr=LR)
print("[INFO] training...")
model.fit({'input': trainData}, {'targets': trainLabels},
validation_set=0.1,
n_epoch=EPOCHS,
snapshot_step=500,
show_metric=False,
run_id=MODEL_NAME)
print("[INFO] evaluating...")
accuracy = model.evaluate(testData, testLabels)[0]
print("[INFO] accuracy: {:.2f}%".format(accuracy * 100))
