def __init__(self, config):
super().__init__()
"""
If only one dataset, then the data set is assumed to be label(there is input_path in CONFIG);
if there are two datasets, then train is assumed to be net input, label
:param config:
:return:
"""
if config.__contains__("input_path"):
self.train_np = None
self.label_np = None
self.input_path = config["input_path"]
self.label_path = self.input_path
self.without_label = True
else:
self.train_np = None
self.label_np = None
self.train_path = config["train_path"]
self.label_path = config["label_path"]
self.without_label = False
# optional
self.batch_size = config["batch_size"] if config.__contains__(
"batch_size") else None
self.shuffle = config["shuffle"] if config.__contains__(
"shuffle") else False
self.processor = DataProcessor()
self.dataloader = Dataloader()
def __init__(self, config):
if config.__contains__("input_path"):
self.test_dnp = None
self.label_dnp = None
self.input_path = config["input_path"]
self.label_path = self.input_path
self.without_label = True
else:
self.test_dnp = None
self.label_dnp = None
self.test_path = config["train_path"]
self.label_path = config["label_path"]
self.without_label = False
self.processor = DataProcessor()
self.dataloader = Dataloader()
def initialization_step(self):
self.curriter=0
## data initialization
if self.net_input_config == "mr":
os.chdir(os.path.dirname(__file__))
mr = sio.loadmat("./mr.mat")["mr"]
self.net_input,_,_ = DataProcessor().normalization_range(mr.reshape(1,1,self.image_shape[1],self.image_shape[0]).transpose(0,1,3,2))# the shape order has been checked!
else:
self.net_input = np.random.randn(1,1,self.image_shape[0],self.image_shape[1])
os.chdir(os.path.dirname(__file__))
self.mask = ~sio.loadmat("./u.mat")["u"].reshape(1,1,self.image_shape[1],self.image_shape[0]).transpose(0,1,3,2).astype(np.bool)
## net initialization
self.initial_net(self.model_cofig)
# x_pre = self.mlem(iters= self.pretrain_iters)
self.net_train(self.x_0,self.pretrain_epoch)
from data.DataProcessor import DataProcessor
from utils.config import config
dp = DataProcessor()
print("Starting augmentation.")
dp.augment_and_save()
print("Finished augmentation. The images were saved to {}".format(
config.AUGMENTED_TRAIN_WALDO))
import numpy as np
from keras.models import load_model
from data.DataProcessor import DataProcessor
from utils.config import config
dp = DataProcessor()
print("Starting precision check.")
# load and preprocess testing data
train_x, train_y, test_x, test_y = dp.load_grayscale()
_, _, test_X, test_Y, _ = dp.preprocess_data(train_x, train_y, test_x, test_y)
model = load_model('version{}_epochs{}_model{}.h5'.format(
config.VERSION, config.NUM_EPOCHS, config.MODEL_VERSION))
preds_org = model.predict(test_X)
preds = np.round(preds_org - config.THRESHOLD + 0.5)
false_pos = 0
false_neg = 0
total_pos = 0
total_neg = 0
for i in range(0, len(preds)):
if test_Y[i][1] == 1:
total_neg += 1
if test_Y[i][0] == 1:
total_pos += 1
if preds[i][0] != test_Y[i][0] or preds[i][1] != test_Y[i][1]:
if test_Y[i][1] == 1:
false_pos += 1
def net_train(self,label,end_epoch):
label,max,min = DataProcessor().normalization_range(label)
output = self.model.trainer.train_DIP(self.model, label, self.net_input, end_epoch,self.mask)
return DataProcessor().normalization_range_reverse(output,max,min)