def build_model(repo='', filename=''):
"""
parser = argparse.ArgumentParser(description='CIFAR 100 Ensemble Prediction')
parser.add_argument('--optimize', type=int, default=0, help='Optimization flag. Set to 1 to perform a randomized '
'search to maximise classification accuracy. \n'
'Set to -1 to get non weighted classification accuracy')
parser.add_argument('--num_tests', type=int, default=20, help='Number of tests to perform when optimizing the '
'ensemble weights for maximizing classification accuracy')
parser.add_argument('--model', type=str, default='wrn', help='Type of model to train')
# Wide ResNet Parameters
parser.add_argument('--wrn_N', type=int, default=2, help='Number of WRN blocks. Computed as N = (n - 4) / 6.')
parser.add_argument('--wrn_k', type=int, default=4, help='Width factor of WRN')
# DenseNet Parameters
parser.add_argument('--dn_depth', type=int, default=40, help='Depth of DenseNet')
parser.add_argument('--dn_growth_rate', type=int, default=12, help='Growth rate of DenseNet')
args = parser.parse_args()
"""
N = 2
k = 4
(trainX, trainY), (testX, testY) = cifar100.load_data()
nb_classes = len(np.unique(testY))
trainX = trainX.astype('float32')
trainX /= 255.0
testX = testX.astype('float32')
testX /= 255.0
trainY = kutils.to_categorical(trainY)
testY_cat = kutils.to_categorical(testY)
if K.image_dim_ordering() == "th":
init = (3, 32, 32)
else:
init = (32, 32, 3)
model = WRN.create_wide_residual_network(init,
nb_classes=100,
N=N,
k=k,
dropout=0.00)
model_prefix = 'WRN-CIFAR100-%d-%d' % (N * 6 + 4, k)
model.compile(loss="categorical_crossentropy",
optimizer="sgd",
metrics=["acc"])
import os
filename = os.path.join(repo, filename)
if os.path.exists(filename):
model.load_weights(filename)
print("Finished compiling")
return model, ((trainX, trainY), (testX, testY))
generator = ImageDataGenerator(rotation_range=15,
width_shift_range=5. / 32,
height_shift_range=5. / 32,
horizontal_flip=False)
generator.fit(trainX, seed=0, augment=True)
if K.image_dim_ordering() == "th":
init = (3, img_rows, img_cols)
else:
init = (img_rows, img_cols, 3)
if model_type == "wrn":
model = WRN.create_wide_residual_network(init,
nb_classes=100,
N=args.wrn_N,
k=args.wrn_k,
dropout=0.00)
model_prefix = 'WRN-CIFAR100-%d-%d' % (args.wrn_N * 6 + 4, args.wrn_k)
else:
model = DN.create_dense_net(nb_classes=100,
img_dim=init,
depth=args.dn_depth,
nb_dense_block=1,
growth_rate=args.dn_growth_rate,
nb_filter=16,
dropout_rate=0.2)
model_prefix = 'DenseNet-CIFAR100-%d-%d' % (args.dn_depth,
args.dn_growth_rate)
def build_model(repo='', filename=''):
parser = argparse.ArgumentParser(
description='CIFAR 10 Ensemble Prediction')
parser.add_argument('--M', type=int, default=5, help='Number of snapshots')
parser.add_argument('--nb_epoch',
type=int,
default=200,
help='Number of training epochs')
parser.add_argument('--alpha_zero',
type=float,
default=0.1,
help='Initial learning rate')
parser.add_argument('--model',
type=str,
default='wrn',
help='Type of model to train')
# Wide ResNet Parameters
parser.add_argument(
'--wrn_N',
type=int,
default=2,
help='Number of WRN blocks. Computed as N = (n - 4) / 6.')
parser.add_argument('--wrn_k',
type=int,
default=4,
help='Width factor of WRN')
# DenseNet Parameters
parser.add_argument('--dn_depth',
type=int,
default=40,
help='Depth of DenseNet')
parser.add_argument('--dn_growth_rate',
type=int,
default=12,
help='Growth rate of DenseNet')
args = parser.parse_args()
''' Snapshot major parameters '''
M = args.M # number of snapshots
nb_epoch = T = args.nb_epoch # number of epochs
alpha_zero = args.alpha_zero # initial learning rate
model_type = str(args.model).lower()
assert model_type in [
'wrn', 'dn'
], 'Model type must be one of "wrn" for Wide ResNets or "dn" for DenseNets'
snapshot = SnapshotCallbackBuilder(T, M, alpha_zero)
batch_size = 128 if model_type == "wrn" else 64
img_rows, img_cols = 32, 32
(trainX, trainY), (testX, testY) = cifar10.load_data()
trainX = trainX.astype('float32')
trainX /= 255.0
testX = testX.astype('float32')
testX /= 255.0
trainY_cat = kutils.to_categorical(trainY)
testY_cat = kutils.to_categorical(testY)
if K.image_dim_ordering() == "th":
init = (3, img_rows, img_cols)
else:
init = (img_rows, img_cols, 3)
model = WRN.create_wide_residual_network(init,
nb_classes=10,
N=args.wrn_N,
k=args.wrn_k,
dropout=0.00)
model_prefix = 'WRN-CIFAR10-%d-%d' % (args.wrn_N * 6 + 4, args.wrn_k)
model.compile(loss="categorical_crossentropy",
optimizer="sgd",
metrics=["acc"])
import os
filename = os.path.join(repo, filename)
if os.path.exists(filename):
model.load_weights(filename)
print("Finished compiling")
return model, ((trainX, trainY_cat), (testX, testY))
