def init_data(mode):
global CLASS_NUM, BATCH_SIZE, inet, cifar_data, data_set, dp, config_name, raw_cifar
assert mode in ["train", "eval"]
CLASS_NUM = settings.config["CLASS_NUM"]
BATCH_SIZE = settings.config["BATCH_SIZE"]
data_set = settings.config["data_set"]
config_name = settings.config["config_name"]
assert data_set in ["cifar10", "svhn", "imagenet"]
data_set = data_set
if data_set == "imagenet":
if mode == "train":
inet = imagenet(BATCH_SIZE, dataset="train")
elif mode == "eval":
inet = imagenet(BATCH_SIZE, dataset="val")
elif data_set == "cifar10":
raw_cifar = cifar10_input.CIFAR10Data("cifar10_data")
if mode == "eval":
cifar_data = raw_cifar.eval_data
elif mode == "train":
cifar_data = raw_cifar.train_data
else:
assert False, "Not implemented"
dp = datapair(CLASS_NUM, BATCH_SIZE)
#LR_DECAY_RATE = 1e-5 # 5e-5
DECAY_STEPS = 1.0
adv_weight = 128
ITER = 100
style_weight = 1
if data_set == "cifar10":
raw_cifar = cifar10_input.CIFAR10Data("cifar10_data")
def get_data(sess):
x_batch, y_batch = raw_cifar.train_data.get_next_batch(
batch_size=BATCH_SIZE, multiple_passes=True)
return x_batch, y_batch
elif data_set == "imagenet":
inet = imagenet(BATCH_SIZE, dataset="train")
def get_data(sess):
x_batch, y_batch = inet.get_next_batch(sess)
return x_batch, y_batch
ENCODER_WEIGHTS_PATH = 'vgg19_normalised.npz'
if data_set == "cifar10":
Decoder_Model = "./trans_pretrained/cifar10transform1.ckpt-574000" #'transform2.ckpt-147000'#./trans_pretrained/transform.ckpt'
elif data_set == "imagenet":
if data_set_name == "imagenet_shallow":
Decoder_Model = "./imagenetshallowtransform1.ckpt.mode2"
elif data_set_name == "imagenet_shallowest":
Decoder_Model = "./imagenetshallowesttransform1.ckpt.mode2"
LEARNING_RATE = 1e-3
LR_DECAY_RATE = 1e-5 # 5e-5
DECAY_STEPS = 1.0
adv_weight = 12 #128
ITER = 1000
style_weight = 1
if data_set == "cifar10":
raw_cifar = cifar10_input.CIFAR10Data("cifar10_data")
def get_data(sess):
x_batch, y_batch = raw_cifar.train_data.get_next_batch(
batch_size=BATCH_SIZE, multiple_passes=True)
return x_batch, y_batch
elif data_set == "imagenet":
inet = imagenet(BATCH_SIZE, "train")
def get_data(sess):
x_batch, y_batch = inet.get_next_batch(sess)
return x_batch, y_batch
ENCODER_WEIGHTS_PATH = 'vgg19_normalised.npz'
if data_set == "cifar10":
# 'transform2.ckpt-147000'#./trans_pretrained/transform.ckpt'
Decoder_Model = "./trans_pretrained/cifar10transform1.ckpt-574000"
elif data_set == "imagenet":
# "./trans_pretrained/imagenetshallowtransform1.ckpt-104000"
Decoder_Model = "./imagenetshallowtransform1.ckpt.mode2"
DECAY_STEPS = 1.0
adv_weight = 128
ITER = 500
CLIP_NORM_VALUE = 10.0
style_weight = 1
if data_set == "cifar10":
raw_cifar = cifar10_input.CIFAR10Data("cifar10_data")
def get_data(sess):
x_batch, y_batch = raw_cifar.eval_data.get_next_batch(
batch_size=BATCH_SIZE, multiple_passes=True)
return x_batch, y_batch
elif data_set == "imagenet":
inet = imagenet(BATCH_SIZE, "val")
def get_data(sess):
x_batch, y_batch = inet.get_next_batch(sess)
return x_batch, y_batch
ENCODER_WEIGHTS_PATH = 'vgg19_normalised.npz'
if data_set == "cifar10":
Decoder_Model = "./trans_pretrained/cifar10transform1.ckpt-574000" #'transform2.ckpt-147000'#./trans_pretrained/transform.ckpt'
elif data_set == "imagenet":
if decoder_name == "imagenet_shallowest":
Decoder_Model = "./imagenetshallowesttransform1.ckpt.mode2"
elif decoder_name == "imagenet_shallow":
Decoder_Model = "./imagenetshallowtransform1.ckpt.mode2" #"./trans_pretrained/imagenetshallowtransform1.ckpt-104000"
