train_index = np.load('train.npy')
valid_index = np.load('valid.npy')
valid_index = valid_index[:24]
b_size = 2
workers = 0
model = model_bbox(b_size, 4)
model.cuda()
optimizer = torch.optim.Adam(model.parameters(), lr=0.001, betas=(0.5, 0.999))
criterion = nn.MSELoss()
load_checkpoint('best_fbb.pth', model, optimizer)
train_set = tumor_dataset(path=train_path,
out_index=train_index,
csv=csv,
read_label=False)
train_loader = DataLoader(train_set,
batch_size=b_size,
shuffle=True,
num_workers=workers)
valid_set = tumor_dataset(path=train_path,
out_index=valid_index,
csv=csv,
read_label=False)
valid_loader = DataLoader(valid_set,
batch_size=b_size,
shuffle=False,
num_workers=workers)
model.eval()
from models import UNet3d_vae
from loss import loss_3d_crossentropy, F1_Loss
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
train_path = '../vae/brats18_data/train_2/'
type1 = ['flair', 't1', 't1ce', 't2']
batch_size = 1
workers = 2
classes = 5
x = 64
y = 64
z = 64
train_set = tumor_dataset(path=train_path)
train_loader = DataLoader(train_set,
batch_size=batch_size,
shuffle=False,
num_workers=workers)
valid_set = tumor_dataset(path=train_path)
valid_loader = DataLoader(valid_set,
batch_size=batch_size,
shuffle=False,
num_workers=workers)
model = UNet3d_vae(4, classes, 64, 64, 64)
print(model)
model.cuda()
optimizer = torch.optim.Adam(model.parameters(),
return gt, self.list[self.out_index[index]]
elif (self.read_label == False):
return feat, self.list[self.out_index[index]]
else:
return feat, gt, self.list[self.out_index[index]]
def __len__(self):
""" Total number of samples in the dataset """
return self.len
b_size = 2
workers = 0
train_set = tumor_dataset(path=train_path,
out_index=train_index,
read_label=False)
train_loader = DataLoader(train_set,
batch_size=b_size,
shuffle=False,
num_workers=workers)
dataloader = train_loader
m = model_bbox(b_size, 4)
m.cuda()
for i, data in enumerate(dataloader):
print(data[0].shape)
img = data[0]
img = img.cuda()
m(img)
return self.len
box_length = 30
def get_yololabel (csv,filename,axis):
min_x = csv.loc[(filename, axis)]['min_x']
max_x = csv.loc[(filename, axis)]['max_x']
min_y = csv.loc[(filename, axis)]['min_y']
max_y = csv.loc[(filename, axis)]['max_y']
x_size = csv.loc[(filename, axis)]['x_size']
y_size = csv.loc[(filename, axis)]['y_size']
new_bbox = np.zeros(8,8,4)
print(min_x,min_y,max_x,max_y)
train_set = tumor_dataset(path = train_path,out_index=train_index)
train_loader = DataLoader(train_set, batch_size=batch_size,shuffle=True, num_workers=workers)
valid_set = tumor_dataset(path = train_path,out_index=valid_index)
valid_loader = DataLoader(valid_set, batch_size=1,shuffle=False, num_workers=workers)
csv = pd.read_csv('tumor_analysis_2d_xy_plane_z_axis.csv' )
csv.set_index(keys=['file_name','z'],inplace=True)
for i , (img,filename) in enumerate(train_loader):
print(img.shape,filename)
break
def main():
b_size = 2
workers = 0
n_epochs = 100
train_set = tumor_dataset(path=train_path,
out_index=train_index,
csv=csv,
read_label=False)
train_loader = DataLoader(train_set,
batch_size=b_size,
shuffle=True,
num_workers=workers)
valid_set = tumor_dataset(path=train_path,
out_index=valid_index,
csv=csv,
read_label=False)
valid_loader = DataLoader(valid_set,
batch_size=b_size,
shuffle=False,
num_workers=workers)
dataloader = train_loader
model = model_bbox(b_size, 4)
model.cuda()
optimizer = torch.optim.Adam(model.parameters(),
lr=0.001,
betas=(0.5, 0.999))
criterion = nn.MSELoss()
total_loss = []
best_loss = np.inf
best_epoch = 0
for epoch in range(n_epochs):
model.train()
epoch_loss = 0
for i, (img, label, _) in enumerate(dataloader):
model.zero_grad()
img = img.cuda()
label = label.float().cuda()
pred = model(img)
loss = criterion(pred, label)
loss.backward()
optimizer.step()
epoch_loss += loss.detach().item()
print('[%d/%d],[%d/%d],loss : %.4f' %
(epoch, n_epochs, i, len(dataloader), loss.detach().item()),
end='\r')
print('\n')
epoch_loss /= len(dataloader)
total_loss.append(epoch_loss)
valid_loss = test(valid_loader)
if (best_loss >= valid_loss):
best_loss = valid_loss
best_epoch = epoch
save_checkpoint('best_fbb.pth', model, optimizer)
print('save best')
save_checkpoint('final_fbb.pth', model, optimizer)
print(
'-----------------------Epoch : %d ,train loss : %.4f ,valid_loss %.4f -------------Best : %d , loss %.4f----------------------'
% (epoch, epoch_loss, valid_loss, best_epoch, best_loss))