def data_gen(data_paths, mask_paths):
"""
get all training images including pet/ct (pet-ct) and mask images
:param data_paths: data paths for PET images
:param mask_paths: paths for mask images
:return: PET images with batch and
"""
no_samples = len(data_paths)
pet_imgs = np.zeros(shape=(no_samples, 1, 96, 96, 96), dtype=np.float32) # change patch shape if necessary
mask_imgs = np.zeros(shape=(no_samples, 1, 96, 96, 96), dtype=np.float32)
for i, (pet_path, mask_path) in tqdm(enumerate(zip(data_paths, mask_paths)), total=no_samples):
pet = sitk.GetArrayFromImage(sitk.ReadImage(pet_path))
mask = sitk.GetArrayFromImage(sitk.ReadImage(mask_path))
# insert one dimension to the existing data as image channel
pet = np.expand_dims(pet, axis=0)
mask = np.expand_dims(mask, axis=0)
# append image
pet_imgs[i] = pet
mask_imgs[i] = mask
# Normalize data and convert label value to either 1 or 0
pet_imgs = normalize(pet_imgs)
mask_imgs = label_converter(mask_imgs)
print("Loading and Process Complete!")
return pet_imgs, mask_imgs
def get_data(data_path):
"""
use (pet_path, ct_path, mask_path) to get corresponding image array
:param data_path: path consists of (pet_path, ct_path, mask_path)
:return: a list of corresponding image path
"""
pet_path, ct_path, mask_path = data_path
pet_img = sitk.GetArrayFromImage(sitk.ReadImage(pet_path))
ct_img = sitk.GetArrayFromImage(sitk.ReadImage(ct_path))
mask = sitk.GetArrayFromImage(sitk.ReadImage(mask_path))
mask = label_converter(mask)
return [pet_img, ct_img, mask]
def data_gen(data_paths, mask_paths):
"""
get all training images including pet/ct (pet-ct) and mask images
:param data_paths: data paths for PET images
:param mask_paths: paths for mask images
:return: PET images with batch and
"""
no_samples = len(data_paths)
imgs = np.zeros(shape=(no_samples, 1, 30, 30, 30),
dtype=np.float32) # change patch shape if necessary
mask_imgs = np.zeros(shape=(no_samples, 1, 30, 30, 30), dtype=np.float32)
for i, (img_path, mask_path) in tqdm(enumerate(zip(data_paths,
mask_paths)),
total=no_samples):
# print(pet_path)
img = sitk.GetArrayFromImage(sitk.ReadImage(img_path))
mask = sitk.GetArrayFromImage(sitk.ReadImage(mask_path))
# insert one dimension to the existing data as image channel
#pet = np.expand_dims(pet, axis=0)
#mask = np.expand_dims(mask, axis=0)
# append image
img = resize(img, (30, 30, 30), mode='constant', preserve_range=True)
mask = resize(mask, (30, 30, 30), mode='constant', preserve_range=True)
imgs[i, 0, :, :, :] = img
mask_imgs[i, 0, :, :] = mask
# Normalize data and convert label value to either 1 or 0
imgs[:, 0, :, :, :] = 3
mask_imgs[:, 0, :, :, :] = 3
imgs = imgs / 255.
mask_imgs = label_converter(mask_imgs)
## AQUI DEBE IR EL CORTE DE LA SECCIÓN DEL CEREBRO
print("Loading and Process Complete!")
return imgs, mask_imgs
from test import test
from dataset import dataset,data_prefetcher
from utils import label_converter,OneHot_encoder
os.environ['CUDA_LAUNCH_BLOCKING'] = "1"
os.environ['CUDA_VISIBLE_DEVICES'] = "4"
opt = {"seed":44,"img_size":(64,84),"max_len":94,"num_epoch":100,"batch_size":80,
"lr":0.1,"grad_clip":5,"model_name":"TransformerBase","train_stage":"1.0",
"previous_stage":None,"split_ratio":0.9,"optimizer":"Adam"
}
with open("./vocab.json","r") as f:
CH_char = json.load(f)
class_num = len(CH_char)+2
converter = label_converter("".join(CH_char))
OneHot = OneHot_encoder(class_num)
def time_interval(second):
s,ms = divmod(int(1000*(second)),1000)
m,s = divmod(s,60)
h,m = divmod(m,60)
return "%02d:%02d:%02d:%03d"%(h,m,s,ms)
def make_model(class_num=class_num,SeqLen=int(opt["img_size"][0]*opt["img_size"][1]/4**2),d_model=512,d_ff=2048,h=8,EncodeLayer_num=2,DecodeLayer_num=1,dropout=0.1):
'''features_extractor'''
features_extractor = ResNet_FeatureExtractor(1,d_model)
"position encoder"
pe = torch.zeros(SeqLen, d_model)
position = torch.arange(0, SeqLen).float().unsqueeze(dim=1)