def train(cycle_num,
dirs,
path_to_net,
plotter,
batch_size=12,
test_split=0.3,
random_state=666,
epochs=100,
learning_rate=0.0001,
momentum=0.9,
num_folds=5,
num_slices=155,
n_classes=4):
"""
Applies training on the network
Args:
cycle_num (int): number of cycle in n-fold (num_folds) cross validation
dirs (string): path to dataset subject directories
path_to_net (string): path to directory where to save network
plotter (callable): visdom plotter
batch_size - default (int): batch size
test_split - default (float): percentage of test split
random_state - default (int): seed for k-fold cross validation
epochs - default (int): number of epochs
learning_rate - default (float): learning rate
momentum - default (float): momentum
num_folds - default (int): number of folds in cross validation
num_slices - default (int): number of slices per volume
n_classes - default (int): number of classes (regions)
"""
print('Setting started', flush=True)
# Creating data indices
# arange len of list of subject dirs
indices = np.arange(len(glob.glob(dirs + '*')))
test_indices, trainset_indices = get_test_indices(indices, test_split)
# kfold index generator
for cv_num, (train_indices, val_indices) in enumerate(
get_train_cv_indices(trainset_indices, num_folds, random_state)):
# splitted the 5-fold CV in 5 jobs
if cv_num != int(cycle_num):
continue
net = U_Net()
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
num_GPU = torch.cuda.device_count()
if num_GPU > 1:
print('Let us use {} GPUs!'.format(num_GPU), flush=True)
net = nn.DataParallel(net)
net.to(device)
criterion = nn.CrossEntropyLoss()
if cycle_num % 2 == 0:
optimizer = optim.SGD(net.parameters(),
lr=learning_rate,
momentum=momentum)
else:
optimizer = optim.Adam(net.parameters(), lr=learning_rate)
scheduler = ReduceLROnPlateau(optimizer, threshold=1e-6, patience=0)
print('cv cycle number: ', cycle_num, flush=True)
start = time.time()
print('Start Train and Val loading', flush=True)
MRIDataset_train = dataset.MRIDataset(dirs, train_indices)
MRIDataset_val = dataset.MRIDataset(dirs, val_indices)
datalengths = {
'train': len(MRIDataset_train),
'val': len(MRIDataset_val)
}
dataloaders = {
'train': get_dataloader(MRIDataset_train, batch_size, num_GPU),
'val': get_dataloader(MRIDataset_val, batch_size, num_GPU)
}
print('Train and Val loading took: ', time.time() - start, flush=True)
# make loss and acc history for train and val separatly
# Setup Metrics
running_metrics_val = runningScore(n_classes)
running_metrics_train = runningScore(n_classes)
val_loss_meter = averageMeter()
train_loss_meter = averageMeter()
itr = 0
iou_best = 0.
for epoch in tqdm(range(epochs), desc='Epochs'):
print('Epoch: ', epoch + 1, flush=True)
phase = 'train'
print('Phase: ', phase, flush=True)
start = time.time()
# Set model to training mode
net.train()
# Iterate over data.
for i, data in tqdm(enumerate(dataloaders[phase]),
desc='Data Iteration ' + phase):
if (i + 1) % 100 == 0:
print('Number of Iteration [{}/{}]'.format(
i + 1, int(datalengths[phase] / batch_size)),
flush=True)
# get the inputs
inputs = data['mri_data'].to(device)
GT = data['seg'].to(device)
subject_slice_path = data['subject_slice_path']
# Clear all accumulated gradients
optimizer.zero_grad()
# Predict classes using inputs from the train set
SR = net(inputs)
# Compute the loss based on the predictions and
# actual segmentation
loss = criterion(SR, GT)
# Backpropagate the loss
loss.backward()
# Adjust parameters according to the computed
# gradients
# -- weight update
optimizer.step()
# Trake and plot metrics and loss, and save network
predictions = SR.data.max(1)[1].cpu().numpy()
GT_cpu = GT.data.cpu().numpy()
running_metrics_train.update(GT_cpu, predictions)
train_loss_meter.update(loss.item(), n=1)
if (i + 1) % 100 == 0:
itr += 1
score, class_iou = running_metrics_train.get_scores()
for k, v in score.items():
plotter.plot(k, 'itr', phase, k, itr, v)
for k, v in class_iou.items():
print('Class {} IoU: {}'.format(k, v), flush=True)
plotter.plot(
str(k) + ' Class IoU', 'itr', phase,
str(k) + ' Class IoU', itr, v)
print('Loss Train', train_loss_meter.avg, flush=True)
plotter.plot('Loss', 'itr', phase, 'Loss Train', itr,
train_loss_meter.avg)
print('Phase {} took {} s for whole {}set!'.format(
phase,
time.time() - start, phase),
flush=True)
# Validation Phase
phase = 'val'
print('Phase: ', phase, flush=True)
start = time.time()
# Set model to evaluation mode
net.eval()
start = time.time()
with torch.no_grad():
# Iterate over data.
for i, data in tqdm(enumerate(dataloaders[phase]),
desc='Data Iteration ' + phase):
if (i + 1) % 100 == 0:
print('Number of Iteration [{}/{}]'.format(
i + 1, int(datalengths[phase] / batch_size)),
flush=True)
# get the inputs
inputs = data['mri_data'].to(device)
GT = data['seg'].to(device)
subject_slice_path = data['subject_slice_path']
# Clear all accumulated gradients
optimizer.zero_grad()
# Predict classes using inputs from the train set
SR = net(inputs)
# Compute the loss based on the predictions and
# actual segmentation
loss = criterion(SR, GT)
# Trake and plot metrics and loss
predictions = SR.data.max(1)[1].cpu().numpy()
GT_cpu = GT.data.cpu().numpy()
running_metrics_val.update(GT_cpu, predictions)
val_loss_meter.update(loss.item(), n=1)
if (i + 1) % 100 == 0:
itr += 1
score, class_iou = running_metrics_val.get_scores()
for k, v in score.items():
plotter.plot(k, 'itr', phase, k, itr, v)
for k, v in class_iou.items():
print('Class {} IoU: {}'.format(k, v), flush=True)
plotter.plot(
str(k) + ' Class IoU', 'itr', phase,
str(k) + ' Class IoU', itr, v)
print('Loss Val', val_loss_meter.avg, flush=True)
plotter.plot('Loss ', 'itr', phase, 'Loss Val', itr,
val_loss_meter.avg)
if (epoch + 1) % 10 == 0:
if score['Mean IoU'] > iou_best:
save_net(path_to_net, batch_size, epoch, cycle_num,
train_indices, val_indices, test_indices, net,
optimizer)
iou_best = score['Mean IoU']
save_output(epoch, path_to_net, subject_slice_path,
SR.data.cpu().numpy(), GT_cpu)
print('Phase {} took {} s for whole {}set!'.format(
phase,
time.time() - start, phase),
flush=True)
# Call the learning rate adjustment function after every epoch
scheduler.step(val_loss_meter.avg)
# save network after training
save_net(path_to_net,
batch_size,
epochs,
cycle_num,
train_indices,
val_indices,
test_indices,
net,
optimizer,
iter_num=None)
class Solver(object):
def __init__(self, config, train_loader, valid_loader, test_loader):
# data loader
self.train_loader = train_loader
self.valid_loader = valid_loader
self.test_loader = test_loader
# Models
self.unet = None
self.optimizer = None
self.img_ch = config['img_ch']
self.output_ch = config['output_ch']
self.criterion = torch.nn.BCELoss() # binary cross entropy loss
# Hyper-parameters
self.lr = config['lr']
self.beta1 = config['beta1'] # momentum1 in Adam
self.beta2 = config['beta2'] # momentum2 in Adam
# Training settings
self.num_epochs = config['num_epochs']
self.num_epochs_decay = config['num_epoches_decay']
self.batch_size = config['batch_size']
# Path
self.model_path = config['model_path']
self.result_path = config['result_path']
self.device = torch.device(
'cuda' if torch.cuda.is_available() else 'cpu')
self.model_type = config['model_type']
self.t = config['t']
self.unet_path = os.path.join(
self.model_path, '%s-%d-%.4f-%d.pkl' %
(self.model_type, self.num_epochs, self.lr, self.num_epochs_decay))
self.best_epoch = 0
self.build_model()
def build_model(self):
"""Build generator and discriminator."""
if self.model_type == 'U_Net':
self.unet = U_Net(img_ch=1, output_ch=1)
elif self.model_type == 'R2U_Net':
self.unet = R2U_Net(img_ch=1, output_ch=1, t=self.t)
#init_weights(self.unet, 'normal')
self.optimizer = optim.Adam(list(self.unet.parameters()), self.lr,
(self.beta1, self.beta2))
self.unet.to(self.device)
def train(self):
"""Print out the network information."""
num_params = 0
for p in self.unet.parameters():
num_params += p.numel(
) # accumulate the number of mmodel parameters
print("The number of parameters: {}".format(num_params))
# ====================================== Training ===========================================#
# network train
if os.path.isfile(self.unet_path):
# Load the pretrained Encoder
self.unet.load_state_dict(torch.load(self.unet_path))
print('%s is Successfully Loaded from %s' %
(self.model_type, self.unet_path))
else:
lr = self.lr
best_unet_score = 0.0
best_epoch = 0
for epoch in range(self.num_epochs):
self.unet.train(True)
epoch_loss = 0
acc = 0. # Accuracy
SE = 0. # Sensitivity (Recall)
SP = 0. # Specificity
PC = 0. # Precision
F1 = 0. # F1 Score
JS = 0. # Jaccard Similarity
DC = 0. # Dice Coefficient
length = 0
for i, (images, GT) in enumerate(self.train_loader):
images, GT = images.to(self.device), GT.to(self.device)
# forward result
SR = self.unet(images)
SR_probs = torch.sigmoid(SR)
SR_flat = SR_probs.view(SR_probs.size(0),
-1) # size(0) is batch_size
GT_flat = GT.view(GT.size(0), -1)
loss = self.criterion(SR_flat, GT_flat)
epoch_loss += loss.item()
# Backprop + optimize
self.unet.zero_grad()
loss.backward()
self.optimizer.step()
acc += get_accuracy(SR, GT)
SE += get_sensitivity(SR, GT)
SP += get_specificity(SR, GT)
PC += get_precision(SR, GT)
F1 += get_F1(SR, GT)
JS += get_JS(SR, GT)
DC += get_DC(SR, GT)
length = length + 1
acc = acc / length
SE = SE / length
SP = SP / length
PC = PC / length
F1 = F1 / length
JS = JS / length
DC = DC / length
# Print the log info
print(
'Epoch [%d/%d], Loss: %.4f, \n[Training] Acc: %.4f, SE: %.4f, SP: %.4f, PC: %.4f,'
' F1: %.4f, JS: %.4f, DC: %.4f' %
(epoch + 1, self.num_epochs, epoch_loss, acc, SE, SP, PC,
F1, JS, DC))
train_accuracy.append(acc)
# Decay learning rate
if (epoch + 1) > (self.num_epochs - self.num_epochs_decay):
lr -= (self.lr / float(self.num_epochs_decay))
for param_group in self.optimizer.param_groups:
param_group['lr'] = lr
print('Decay learning rate to lr: {}.'.format(lr))
# ===================================== Validation ====================================#
self.unet.train(False)
self.unet.eval()
acc = 0. # Accuracy
SE = 0. # Sensitivity (Recall)
SP = 0. # Specificity
PC = 0. # Precision
F1 = 0. # F1 Score
JS = 0. # Jaccard Similarity
DC = 0. # Dice Coefficient
length = 0
for i, (images, GT) in enumerate(self.valid_loader):
images, GT = images.to(self.device), GT.to(self.device)
SR = torch.sigmoid(self.unet(images))
acc += get_accuracy(SR, GT)
SE += get_sensitivity(SR, GT)
SP += get_specificity(SR, GT)
PC += get_precision(SR, GT)
F1 += get_F1(SR, GT)
JS += get_JS(SR, GT)
DC += get_DC(SR, GT)
length = length + 1
acc = acc / length
SE = SE / length
SP = SP / length
PC = PC / length
F1 = F1 / length
JS = JS / length
DC = DC / length
unet_score = JS + DC
print('[Validation] Acc: %.4f, SE: %.4f, SP: %.4f, PC: %.4f, '
'F1: %.4f, JS: %.4f, DC: %.4f' %
(acc, SE, SP, PC, F1, JS, DC))
validation_accuracy.append(acc)
if unet_score > best_unet_score:
best_unet_score = unet_score
self.best_epoch = epoch
best_unet = self.unet.state_dict(
) # contain best parameters for each layer
print('Best %s model score : %.4f' %
(self.model_type, best_unet_score))
torch.save(best_unet, self.unet_path)
def test(self):
self.unet.load_state_dict(torch.load(self.unet_path))
self.unet.eval()
acc = 0. # Accuracy
SE = 0. # Sensitivity (Recall)
SP = 0. # Specificity
PC = 0. # Precision
F1 = 0. # F1 Score
JS = 0. # Jaccard Similarity
DC = 0. # Dice Coefficient
length = 0
result = []
for i, (images, GT) in enumerate(self.test_loader):
images = images.to(self.device)
GT = GT.to(self.device)
SR = torch.sigmoid(self.unet(images))
acc += get_accuracy(SR, GT)
SE += get_sensitivity(SR, GT)
SP += get_specificity(SR, GT)
PC += get_precision(SR, GT)
F1 += get_F1(SR, GT)
JS += get_JS(SR, GT)
DC += get_DC(SR, GT)
length = length + 1
SR = SR.to('cpu')
SR = SR.detach().numpy()
result.extend(SR)
acc = acc / length
SE = SE / length
SP = SP / length
PC = PC / length
F1 = F1 / length
JS = JS / length
DC = DC / length
unet_score = JS + DC
reconstruct_image(self, np.array(result))
f = open(os.path.join(self.result_path, 'result.csv'),
'a',
encoding='utf-8',
newline='')
wr = csv.writer(f)
wr.writerow([
self.model_type,
acc,
SE,
SP,
PC,
F1,
JS,
DC,
self.lr,
self.best_epoch,
self.num_epochs,
self.num_epochs_decay,
])
f.close()