def get_model(N, k, nb_classes):
model_script_path = os.path.join(os.path.dirname(__file__), '..')
model_script_path = os.path.abspath(model_script_path)
sys.path.append(model_script_path)
import wide_residual_network as wrn
init_shape = (32, 32, 3)
model = wrn.create_wide_residual_network(init_shape,
nb_classes=nb_classes,
N=N,
k=k,
dropout=0.00)
return model
generator = ImageDataGenerator(rotation_range=10,
width_shift_range=5. / 32,
height_shift_range=5. / 32,
horizontal_flip=True)
generator.fit(trainX, seed=0, augment=True)
init_shape = (3, 32, 32) if K.image_dim_ordering() == 'th' else (32, 32, 3)
# For WRN-16-8 put N = 2, k = 8
# For WRN-28-10 put N = 4, k = 10
# For WRN-40-4 put N = 6, k = 4
model = wrn.create_wide_residual_network(init_shape,
nb_classes=10,
N=4,
k=8,
dropout=0.0)
model.summary()
#plot_model(model, "WRN-28-8.png", show_shapes=False)
model.compile(loss="categorical_crossentropy",
optimizer="adam",
metrics=["acc"])
print("Finished compiling")
print("Allocating GPU memory")
# model.load_weights("weights/WRN-28-8 Weights.h5")
# print("Model loaded.")
# color preprocessing
x_train45, x_val, y_train45, y_val = train_test_split(
x_train, y_train, test_size=0.1,
random_state=seed) # random_state = seed
x_train45, x_val, x_test = color_preprocessing(x_train45, x_val, x_test)
y_train45 = keras.utils.to_categorical(y_train45, num_classes)
y_val = keras.utils.to_categorical(y_val, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)
# build network
img_input = Input(shape=(img_rows, img_cols, img_channels))
model = wrn.create_wide_residual_network(img_input,
nb_classes=num_classes,
N=n,
k=growth_rate,
dropout=0.0)
print(model.summary())
# set optimizer
sgd = optimizers.SGD(lr=.1, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy',
optimizer=sgd,
metrics=['accuracy'])
# set callback
change_lr = LearningRateScheduler(scheduler)
checkpointer = ModelCheckpoint('model_wide_28_10_c10_best.hdf5',
verbose=1,
save_best_only=True)
cbks = [change_lr, checkpointer]
# Set up the callbacks
metrics_callback = MetricsCallback(validation_data=(X_test, Y_test))
checkpoint_callback = CheckpointCallback(monitor='accuracy',
verbose=verbose,
save_weights_only=False)
#lr_callback = LearningRateScheduler(schedule=schedule)
callbacks = [metrics_callback, checkpoint_callback]
#For WRN-28-10 N = (nb_layers-4)/6 = (28-4)/6 = 4, k=10 (maybe memory constraints!!!) 4.17%
#For WRN-16-8 N = (nb_layers-4)/6 = (16-4)/6 = 2, k=8 4.81%
#For WRN-40-4 N = (nb_layers-4)/6 = (16-4)/6 = 2, k=8 4.97%
model = wrn.create_wide_residual_network(img_shape,
nb_classes=8,
N=4,
k=10,
dropout=0.5)
# set the optimizer
sgd = SGD(lr=0.1, decay=1e-5, momentum=0.9, nesterov=True)
# Compile the model
model.compile(loss='categorical_crossentropy', optimizer=sgd)
#model.load_weights('../logs_nodrop_128_zca/model-epoch-000032-acc_test-0.86199.hdf5')
# Model Summary
print(model.summary())
model.fit_generator(generator.flow(X_train, Y_train, batch_size=batch_size),
wtype, embed_dim, scale, N, k, batch_size, epoch_num, target_blk_id)
modelname_prefix = os.path.join(
RESULT_PATH, 'wrn_' + history_hdf_path.replace('/', '_'))
# Initialize the process for the watermark
b = np.ones((1, embed_dim))
watermark_regularizer = WatermarkRegularizer(scale,
b,
wtype=wtype,
randseed=randseed)
init_shape = (3, 32, 32) if K.image_dim_ordering() == 'th' else (32, 32, 3)
model = wrn.create_wide_residual_network(
init_shape,
nb_classes=classes_num,
N=N,
k=k,
dropout=0.00,
wmark_regularizer=watermark_regularizer,
target_blk_num=target_blk_id)
model.summary()
print('Watermark matrix:\n{}'.format(watermark_regularizer.get_matrix()))
# Training process
sgd = SGD(lr=0.1, momentum=0.9, nesterov=True)
model.compile(loss="categorical_crossentropy",
optimizer=sgd,
metrics=["acc"])
if len(base_modelw_fname) > 0:
model.load_weights(base_modelw_fname)
print("Finished compiling.")
else:
return 0.0008
lrschedule = LearningRateScheduler(lr_scheduler_fxn, verbose=1)
tboard = TensorBoard(log_dir=os.path.join(root_folder, "logs"))
checkpoint = ModelCheckpoint(
filepath=os.path.join(root_folder, "checkpoints",
"wts_{epoch:02d}-{loss:.2f}.h5"),
verbose=True,
save_weights_only=True
)
# Network
model = wrn.create_wide_residual_network(
(32,32,3),
nb_classes=10,
N=4, k=10,
dropout=0.0
)
print(model.summary())
sgd = SGD(lr=0.1, momentum=0.9, nesterov=True)
model.compile(loss="categorical_crossentropy", optimizer=sgd, metrics=["accuracy"])
# Train
history = model.fit_generator(
train_dgen.flow(trainX, trainY, batch_size=batch_size),
steps_per_epoch=len(trainX) // batch_size + 1,
validation_data=test_dgen.flow(testX, testY, batch_size=batch_size),
validation_steps=len(testX) // batch_size + 1,
epochs=epochs,
callbacks=[lrschedule, tboard, checkpoint]
)
import time
if lib.isnotebook():
from keras_tqdm import TQDMNotebookCallback as KerasCallBack
else:
from keras_tqdm import TQDMCallback as KerasCallBack
# In[5]:
print('Building model...')
K.set_image_data_format('channels_last')
# wrn 28 10
model = wrn.create_wide_residual_network((32, 32, 3), nb_classes=100, N=5, k=10, dropout=0.3, verbose=1)
# In[6]:
print('Loading data...')
data = lib.load_data()
x_train = data['x_train']
y_train = data['y_train']
# In[16]:
x_train.shape, y_train.shape
fill_mode="constant",
cval = 0
)
img_gen.fit(X_train45)
Y_train45 = np_utils.to_categorical(Y_train45, nb_classes) # 1-hot vector
y_val = np_utils.to_categorical(y_val, nb_classes)
y_test = np_utils.to_categorical(y_test, nb_classes)
init_shape = (3, 32, 32) if K.image_dim_ordering() == 'th' else (32, 32, 3)
# For WRN-16-8 put N = 2, k = 8
# For WRN-28-10 put N = 4, k = 10
# For WRN-40-4 put N = 6, k = 4
model = wrn.create_wide_residual_network(init_shape, nb_classes=nb_classes, N=2, k=4, dropout=0.00)
model.summary()
#plot_model(model, "WRN-16-8.png", show_shapes=False)
model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["acc"])
print("Finished compiling")
#model.load_weights("weights/WRN-16-8 Weights.h5")
print("Model loaded.")
hist = model.fit_generator(img_gen.flow(X_train45, Y_train45, batch_size=batch_size, shuffle=True),
steps_per_epoch=len(X_train45) // batch_size, epochs=nb_epoch,
callbacks=[callbacks.ModelCheckpoint("WRN_16_4_Weights_cifar100.h5",
monitor="val_acc",
save_best_only=True,
testX = testX.astype('float32')
testX = (testX - testX.mean(axis=0)) / (testX.std(axis=0))
trainY = kutils.to_categorical(trainY)
testY = kutils.to_categorical(testY)
generator = ImageDataGenerator(rotation_range=10,
width_shift_range=5./32,
height_shift_range=5./32,)
init_shape = (3, 32, 32) if K.image_dim_ordering() == 'th' else (32, 32, 3)
# For WRN-16-8 put N = 2, k = 8
# For WRN-28-10 put N = 4, k = 10
# For WRN-40-4 put N = 6, k = 4
model = wrn.create_wide_residual_network(init_shape, nb_classes=10, N=2, k=8, dropout=0.00)
model.summary()
plot_model(model, "WRN-16-8.png", show_shapes=False)
model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["acc"])
print("Finished compiling")
#model.load_weights("weights/WRN-16-8 Weights.h5")
print("Model loaded.")
model.fit_generator(generator.flow(trainX, trainY, batch_size=batch_size), steps_per_epoch=len(trainX) // batch_size, epochs=nb_epoch,
callbacks=[callbacks.ModelCheckpoint("weights/WRN-16-8 Weights.h5",
monitor="val_acc",
save_best_only=True,
verbose=1)],
width_shift_range=4.0 /
32, # randomly shift images horizontally (fraction of total width)
height_shift_range=4.0 /
32, # randomly shift images vertically (fraction of total height)
fill_mode='constant',
cval=0,
horizontal_flip=True, # randomly flip images
# vertical_flip=False
)
# model = wrn.build_model((32, 32, 3,), classes=100, n=4, k=10, dropout=0.3, weight_decay=0.0005, verbose=True)
model = wrn2.create_wide_residual_network((
32,
32,
3,
),
nb_classes=num_classes,
N=4,
k=10,
dropout=0.3)
model.compile(loss='categorical_crossentropy',
optimizer=optimizers.SGD(lr=0.1,
momentum=0.9,
decay=0.0,
nesterov=True),
metrics=['accuracy'])
model.summary()
train_data_generator.fit(train_data)
def __init__(self, weightsFile):
(self.trainX, trainY), (testX, testY) = cifar10.load_data()
init_shape = (3, 32, 32) if K.image_dim_ordering() == 'th' else (32, 32, 3)
self.model = wrn.create_wide_residual_network(init_shape, nb_classes=10, N=2, k=8, dropout=0.00)
self.model.load_weights(weightsFile)
self.graph = tf.get_default_graph()
generator = ImageDataGenerator(rotation_range=10,
width_shift_range=5. / 32,
height_shift_range=5. / 32,
horizontal_flip=True)
generator.fit(trainX, seed=0, augment=True)
init = Input(shape=(3, img_rows, img_cols), )
# For WRN-16-8 put N = 2, k = 8
# For WRN-28-10 put N = 4, k = 10
# For WRN-40-4 put N = 6, k = 4
wrn_16_8 = wrn.create_wide_residual_network(init,
nb_classes=10,
N=2,
k=8,
dropout=0.00)
model = Model(input=init, output=wrn_16_8)
model.summary()
#plot(model, "WRN-16-8.png", show_shapes=False)
model.compile(loss="categorical_crossentropy",
optimizer="adam",
metrics=["acc"])
print("Finished compiling")
model.load_weights("weights/WRN-16-8 Weights.h5")
print("Model loaded.")