def __init__(self):
# ID and Name
self.id = 109
self.experiment_name = "pancreas_2D_utrans_{}".format(self.id)
self.debug = False
# System
self.checkpointsBasePath = "./checkpoints/"
self.checkpointsBasePathMod = self.checkpointsBasePath + self.experiment_name + '/'
self.datapath = "/local/DEEPLEARNING/TCIA/"
self.split = 0
# GPU
self.gpu = '0'
os.environ["CUDA_VISIBLE_DEVICES"] = self.gpu
# Model
self.channels = [64, 128, 256, 512, 1024]
self.channels = [int(x) for x in self.channels]
self.net = UNetTransformer(filters=self.channels,
n_classes=2,
in_channels=1,
n_heads=1,
dim='2d',
bn=True,
up_mode='deconv')
self.n_parameters = count_parameters(self.net)
print("N PARAMS : {}".format(self.n_parameters))
self.model_path = './checkpoints/models/pancreas2D_utrans_kaiming_deconv.pth'
self.load_model()
self.n_classes = 2
# self.transform = tf.Compose([
# tf.RandomAffine(degrees = 10,
# scale = (0.9,1.1),
# translate = (0.1, 0.1)),
# tf.ToTensor(),
# ])
self.transform = True
# Training
self.start_epoch = 0
self.epoch = 25
self.loss = torch.nn.CrossEntropyLoss()
self.batchsize = 3
self.lr_rate = 1e-4
self.final_lr_rate = 1e-5
self.optimizer = optim.Adam(self.net.parameters(), lr=self.lr_rate)
self.optimizer.zero_grad()
self.validate_every_k_epochs = 1
# self.lr_scheduler = get_scheduler(self.optimizer, "constant", self.lr_rate)
# self.lr_scheduler = optim.lr_scheduler.ExponentialLR(self.optimizer, 0.99)
self.decay = (self.lr_rate / self.final_lr_rate - 1) / self.epoch
self.lr_scheduler = get_scheduler(self.optimizer, "po", self.lr_rate,
self.decay)
# Other
self.classes_name = ['background', 'pancreas']
class ExpConfig():
def __init__(self):
# ID and Name
self.id = 102
self.experiment_name = "pancreas_2D_utrans_{}".format(self.id)
self.debug = False
# System
self.checkpointsBasePath = "./checkpoints/"
self.checkpointsBasePathMod = self.checkpointsBasePath + self.experiment_name + '/'
self.datapath = "/local/DEEPLEARNING/TCIA/"
self.split = 0
# GPU
self.gpu = '1'
os.environ["CUDA_VISIBLE_DEVICES"] = self.gpu
# Model
self.channels = [64, 128, 256, 512, 1024]
self.channels = [int(x) for x in self.channels]
self.net = UNetTransformer(filters=self.channels,
n_classes=2,
in_channels=1,
n_heads=1,
dim='2d',
bn=False)
self.n_parameters = count_parameters(self.net)
print("N PARAMS : {}".format(self.n_parameters))
self.model_path = './checkpoints/models/pancreas2D_utrans_kaiming_nobn.pth'
self.load_model()
self.n_classes = 2
self.transform = tf.Compose([
tf.RandomAffine(degrees=10, scale=(0.9, 1.1),
translate=(0.1, 0.1)),
tf.ToTensor(),
])
# Training
self.start_epoch = 0
self.epoch = 20
self.loss = torch.nn.CrossEntropyLoss()
self.batchsize = 4
self.lr_rate = 1e-4
self.optimizer = optim.Adam(self.net.parameters(), lr=self.lr_rate)
self.optimizer.zero_grad()
self.validate_every_k_epochs = 1
# self.lr_scheduler = get_scheduler(self.optimizer, "constant", self.lr_rate)
self.lr_scheduler = optim.lr_scheduler.ExponentialLR(
self.optimizer, 0.99)
# Other
self.classes_name = ['background', 'pancreas']
def set_data(self, split=0):
# Data
trainDataset = SplitTCIA2DDataset(self.datapath,
self.split,
self.generate_splits('train'),
transform=self.transform,
mode='train')
validDataset = SplitTCIA2DDataset(self.datapath,
self.split,
self.generate_splits('test'),
mode='valid')
self.trainDataLoader = DataLoader(dataset=trainDataset,
num_workers=1,
batch_size=self.batchsize,
shuffle=True)
self.valDataLoader = DataLoader(dataset=validDataset,
num_workers=1,
batch_size=self.batchsize,
shuffle=False)
def generate_splits(self, sets, i=0):
splt_1 = [
19, 23, 55, 30, 60, 26, 12, 6, 64, 38, 42, 62, 69, 5, 14, 63, 31
]
splt_2 = [
46, 59, 81, 61, 41, 15, 49, 43, 48, 32, 11, 37, 53, 17, 47, 21, 2
]
splt_3 = [67, 56, 51, 35, 66, 10, 39, 76, 58, 7, 44, 40, 70, 65, 1, 29]
splt_4 = [27, 52, 33, 45, 13, 72, 80, 8, 82, 73, 34, 16, 18, 71, 9, 28]
splt_5 = [57, 22, 68, 78, 74, 54, 20, 24, 4, 79, 36, 75, 25, 50, 3, 77]
data_splits = {'train': [], 'test': []}
all_splits = [splt_1, splt_2, splt_3, splt_4, splt_5]
data_splits['test'] = all_splits[self.split]
data_splits['train'] = []
for k in [i for i in all_splits[:self.split]
] + all_splits[self.split + 1:]:
data_splits['train'] += k
return data_splits[sets]
def load_model(self):
print('LOAD MODEL ...')
if not os.path.exists(self.model_path):
torch.save(self.net.state_dict(), self.model_path)
else:
self.net.load_state_dict(torch.load(self.model_path))
def net_stats(self):
s = 0
for p in self.net.parameters():
if p.requires_grad:
s += p.sum()
print('Mean :', s.item() / self.n_parameters)