def schedule(dataset, lr):
if dataset == 'cifar10':
steps = [82, 123]
lrs = [lr, lr / 10, lr / 100]
else:
raise ValueError('Schedule not defined for dataset: %s' % dataset)
return Step(steps, lrs)
def schedule(dataset, lr):
if dataset == 'mnist':
steps = [20]
lrs = [lr, lr / 10]
else:
raise ValueError('Schedule not defined for dataset: %s.' % dataset)
return Step(steps, lrs)
def schedule(dataset, lr):
if dataset in ['cifar10', 'cifar100']:
steps = [25 * i for i in range(1, 10)]
lrs = [lr / 2**i for i in range(10)]
elif dataset == 'svhn':
steps = [20, 30]
lrs = [lr, lr / 10, lr / 100]
else:
raise ValueError('Schedule not defined for dataset: %s' % dataset)
return Step(steps, lrs)
model = DenseNet(10, img_dim, depth, nb_dense_block, growth_rate,
nb_filter, dropout_rate, weight_decay)
opt = SGD(lr=learning_rate, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy',
optimizer=opt,
metrics=["accuracy"])
offset = 0
save_path = os.path.join('./log', 'densenet')
callbacks = []
callbacks.append(ModelCheckpoint(save_path + '.h5'))
steps = [nb_epoch / 2 - offset, 3 * nb_epoch / 4 - offset]
schedule = Step(steps,
[learning_rate, 0.1 * learning_rate, 0.01 * learning_rate],
verbose=1)
callbacks.append(schedule)
hist1 = model.fit(X_train,
Y_train,
batch_size=batch_size,
epochs=nb_epoch,
verbose=0,
validation_data=(X_test, Y_test),
callbacks=callbacks,
shuffle=True)
with open('./log/cifar10_hist', 'wb') as file:
pickle.dump(hist1.history, file)
train_set, _, test_set = cifar10.load('data/cifar10',
valid_ratio=0.,
gcn=False,
zca=False)
train_set['data'] = preprocess_data(train_set['data'])
test_set['data'] = preprocess_data(test_set['data'])
print 'Data loaded.'
# Train for 300 epochs as in the paper
nb_epoch = 300
offset = 0
# Define callbacks for checkpointing and scheduling
callbacks = []
callbacks.append(ModelCheckpoint(save_path + '.h5'))
steps = [nb_epoch / 2 - offset, 3 * nb_epoch / 4 - offset]
schedule = Step(steps, [0.1, 0.01, 0.001], verbose=1)
callbacks.append(schedule)
# Custom meta checkpoint that saves training information at each epoch.
callbacks.append(MetaCheckpoint(save_path + '.meta', schedule=schedule))
if augmented:
data_gen = ImageDataGenerator(horizontal_flip=True,
width_shift_range=0.125,
height_shift_range=0.125,
fill_mode='constant')
data_iter = data_gen.flow(train_set['data'],
train_set['labels'],
batch_size=64,
shuffle=True)
model.fit_generator(data_iter,
valid_ratio=1. / 6,
normalize=True)
# Load our model. This is a simple model taken from tensorflow's tutorial
model = simple_model()
# Use the adam optimizer, since we are using a schedule callback. The
# learning rate we define here doesn't really matter
adam = optimizers.Adam(lr=1e-4)
model.compile(optimizer=adam,
loss='categorical_crossentropy',
metrics=['accuracy'])
# Callback that implements learning rate schedule
schedule = Step([20], [1e-4, 1e-5])
history = model.fit(train_set['data'],
train_set['labels'],
batch_size=128,
nb_epoch=40,
validation_data=(valid_set['data'], valid_set['labels']),
callbacks=[schedule],
verbose=2,
shuffle=True)
(loss, acc) = model.evaluate(test_set['data'],
test_set['labels'],
batch_size=128)
print('Test set loss = %g, accuracy = %g' % (loss, acc))