def open_rgby(path, filename, train=True, stats=None):
""" a function that reads RGBY image """
#colors = ['red', 'green', 'blue', 'yellow']
colors = ['red', 'green', 'blue']
flags = cv2.IMREAD_GRAYSCALE
img = [
cv2.imread(os.path.join(path, filename + '_' + color + '.png'),
flags).astype(np.float32) for color in colors
]
img = [cv2.resize(x, (512, 512)) / 255 for x in img]
img = np.stack(img, axis=-1)
##### do not normalize
stats = None
if not stats is None:
m, s = stats
img = transforms.Normalize(img, m, s)
if train:
img = transforms.RandomRotate(img, 30)
img = transforms.RandomDihedral(img)
img = transforms.RandomLighting(img, 0.05, 0.05)
return img
def main():
args = parser.parse_args()
step = 0
exp_name = f'{args.name}_{hp.max_lr}_{hp.cycle_length}'
transforms = segtrans.JointCompose([segtrans.Resize(400),
segtrans.RandomRotate(0, 90),
segtrans.RandomCrop(256, 256),
segtrans.ToTensor(),
segtrans.Normalize(mean=hp.mean,
std=hp.std)])
val_transforms = segtrans.JointCompose([segtrans.PadToFactor(),
segtrans.ToTensor(),
segtrans.Normalize(mean=hp.mean,
std=hp.std)])
train_dataset = DSBDataset(f'{args.data}/train', transforms=transforms)
val_dataset = DSBDataset(f'{args.data}/val', transforms=val_transforms)
model = Unet()
if args.checkpoint:
checkpoint = torch.load(args.checkpoint)
model.load_state_dict(checkpoint['state'])
step = checkpoint['step']
exp_name = checkpoint['exp_name']
optimizer = Adam(model.parameters(), lr=hp.max_lr)
if args.find_lr:
scheduler = LRFinderScheduler(optimizer)
else:
scheduler = SGDRScheduler(optimizer, min_lr=hp.min_lr,
max_lr=hp.max_lr, cycle_length=hp.cycle_length, current_step=step)
model.cuda(device=args.device)
train(model, optimizer, scheduler, train_dataset, val_dataset,
n_epochs=args.epochs, batch_size=args.batch_size,
exp_name=exp_name, device=args.device, step=step)
def train(model, criterion, optimizer, lr_scheduler, data_dir, lr):
"""
docstring
Parameters:
-----------
model: pytorch model
criterion:
optimizer:
lr_scheduler:
data_dir:
"""
time_at_start = time.time()
best_acc = 0.0
history = {
"train_acc": [],
"train_loss": [],
"test_acc": [],
"test_loss": []
}
datasets = make_datasets(data_dir, transforms=transforms.RandomRotate())
dataloader = make_dataloaders(datasets)
for epoch in range(NUM_EPOCHS):
print("Epoch {}/{}".format(epoch, NUM_EPOCHS - 1))
print("=" * 10)
for phase in ["train", "test"]:
if phase == "train":
optimizer = lr_scheduler(optimizer, epoch, lr)
model.train(True)
else:
model.train(False)
print("Setting model to {} mode".format(phase))
running_loss = 0.0
running_corrects = 0
len_data = len(datasets[phase])
for index, data in enumerate(dataloader[phase]):
inputs, labels, _ = data
if USE_GPU:
inputs = torch.autograd.Variable(inputs.cuda())
labels = torch.autograd.Variable(labels.cuda())
else:
inputs = torch.autograd.Variable(inputs)
labels = torch.autograd.Variable(labels)
# zero parameter gradients before the forward pass
optimizer.zero_grad()
outputs = model(inputs)
_, preds = torch.max(outputs.data, 1)
labels = labels.view(-1)
loss = criterion(outputs, labels)
if phase == "train":
loss.backward()
optimizer.step()
running_loss += loss.data[0]
running_corrects += torch.sum(preds == labels.data)
if VERBOSE:
print_model_stats(index, BATCH_SIZE, len_data,
loss.data[0])
print("\n")
# epoch stats
epoch_loss = running_loss / len_data
epoch_acc = running_corrects / len_data
history["{}_acc".format(phase)].append(epoch_acc)
history["{}_loss".format(phase)].append(epoch_loss)
print("{} Loss: {:.4f} | Acc: {:.4f}".format(
phase, epoch_loss, epoch_acc))
# checkpoint the best model based on validation accuracy
if phase == "test" and epoch_acc > best_acc:
best_acc = epoch_acc
model_path = "{}_checkpoint".format(SAVE_PATH)
print("checkpointing model at {}".format(model_path))
torch.save(model.state_dict(), model_path)
# write history to JSON file
history_path = "{}_history".format(SAVE_PATH)
with open(history_path, "w") as f:
json.dump(history, f, indent=4)
time_elapsed = time.time() - time_at_start
print("Training complete in {:.0f}m {:.0f}s".format(
time_elapsed // 60, time_elapsed % 60))
print("Best validation accuracy: {:.4f}".format(best_acc))
def train_model(model, criterion, optimizer, lr_scheduler):
"""docstring"""
history = {"train_acc": [], "train_loss": [],
"test_acc": [], "test_loss": []}
transform = transforms.RandomRotate()
datasets = {}
train_data = dataset.CSVDataset(
data_dir=DATA_DIR, csv=TRAIN_CSV,
complete_csv=DATAFRAME, transforms=transform
)
train_data.equalise_groups("MoA", under_sample=False)
datasets["train"] = train_data
datasets["test"] = dataset.CSVDataset(
data_dir=DATA_DIR, csv=TEST_CSV, complete_csv=DATAFRAME
)
dataloader = {}
dataloader["train"] = torch.utils.data.DataLoader(
datasets["train"], batch_size=BATCH_SIZE,
shuffle=True, num_workers=NUM_WORKERS, pin_memory=True
)
dataloader["test"] = torch.utils.data.DataLoader(
datasets["test"], batch_size=BATCH_SIZE,
shuffle=False, num_workers=NUM_WORKERS, pin_memory=True
)
for epoch in range(NUM_EPOCHS):
print("Epoch {}/{}".format(epoch, NUM_EPOCHS-1))
print("="*10)
# each epoch has training and validation phases
optimizer = lr_scheduler(optimizer, epoch)
model.train(True)
running_loss = 0.0
running_corrects = 0
len_data = len(datasets["train"])
for data in tqdm(dataloader["train"]):
# ignore parent_img labels during training, these are only needed in testing
inputs, labels, _ = data
if USE_GPU:
inputs = torch.autograd.Variable(inputs).cuda()
labels = torch.autograd.Variable(labels).cuda()
else:
inputs = torch.autograd.Variable(inputs)
labels = torch.autograd.Variable(labels)
# zero the parameter gradients before the forward pass
optimizer.zero_grad()
# forward pass
outputs = model(inputs)
_, preds = torch.max(outputs.data, 1)
labels = labels.view(-1)
loss = criterion(outputs, labels)
# backprop if in the training phase
loss.backward()
optimizer.step()
running_loss += loss.data[0]
running_corrects += torch.sum(preds == labels.data)
print("\n")
# epoch stats for train and validation phases
epoch_loss = running_loss / len_data
epoch_acc = running_corrects / len_data
history["train_acc"].append(epoch_acc)
history["train_loss"].append(epoch_loss)
print("Loss: {:.4f} | Acc: {:.4f}".format(epoch_loss, epoch_acc))
# write history dict as a JSON file at each epoch
history_path = "{}_history".format(SAVE_PATH)
with open(history_path, "w") as f:
json.dump(history, f, indent=4)
# Convert model to test mode and make predictions on test set
# these need to be recorded so that the aggregate prediction
# on all the cells in an image can be calculated.
#
# This can be done by grouping on the 'img_id' column in the .csv file
print("\n")
print("=" * 10)
print("Testing")
print("=" * 10)
model.eval()
parent_imgs = []
predictions_list = []
actual_vals = []
for data in tqdm(dataloader["test"]):
test_inputs, test_labels, parent_img = data
test_labels = list(test_labels.cpu().numpy())
if USE_GPU:
test_inputs = torch.autograd.Variable(test_inputs).cuda()
test_outputs = model(test_inputs)
_, predictions = torch.max(test_outputs.data, 1)
predictions = list(predictions.cpu().numpy())
parent_imgs.append(parent_img)
predictions_list.extend(predictions)
actual_vals.extend(test_labels)
for i, j, k in zip(parent_imgs, predictions_list, actual_vals):
print(i, j, k)
return image, label
def __len__(self):
assert self.trainDataSize == self.maskDataSize
assert self.trainDataSize == self.dataBoxSize
return self.trainDataSize
if __name__ == '__main__':
from torch.utils.data import DataLoader
transforms = [
# Transforms.RandomCrop(2300, 2300),
Transforms.RondomFlip(),
Transforms.RandomRotate(15),
Transforms.Log(0.5),
Transforms.Blur(0.2),
Transforms.ToTensor(),
Transforms.ToGray()
]
dataset = UNetDataset('./data/train',
'./data/train_cleaned',
transform=transforms)
dataLoader = DataLoader(dataset=dataset,
batch_size=32,
shuffle=True,
num_workers=0)
for index, (batch_x, batch_y) in enumerate(dataLoader):
print(batch_x.size(),
def train():
# dataset
transforms = [
Transforms.ToGray(),
Transforms.RondomFlip(),
Transforms.RandomRotate(15),
Transforms.RandomCrop(48, 48),
Transforms.Log(0.5),
# Transforms.EqualizeHist(0.5),
# Transforms.Blur(0.2),
Transforms.ToTensor()
]
dataset = UNetDataset('./data/train/',
'./data/train_cleaned/',
transform=transforms)
dataLoader = DataLoader(dataset=dataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=0)
# init model
net = UNet(1, 2).to(device)
optimizer = optim.Adam(net.parameters(), lr=LR)
loss_func = nn.CrossEntropyLoss().to(device)
# load weight
if os.path.exists(weight_with_optimizer):
checkpoint = torch.load(weight_with_optimizer)
net.load_state_dict(checkpoint['net'])
optimizer.load_state_dict(checkpoint['optimizer'])
print('load weight')
# train
for epoch in range(EPOCH):
# train
for step, (batch_x, batch_y) in enumerate(dataLoader):
# import cv2
# import numpy as np
# display = np.concatenate(
# (batch_x[0][0].numpy(), batch_y[0][0].numpy().astype(np.float32)),
# axis=1
# )
# cv2.imshow('display', display)
# cv2.waitKey()
batch_x = batch_x.to(device)
batch_y = batch_y.squeeze(1).to(device)
output = net(batch_x)
loss = loss_func(output, batch_y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
print('epoch: %d | loss: %.4f' % (epoch, loss.data.cpu()))
# save weight
if (epoch + 1) % 10 == 0:
torch.save(
{
'net': net.state_dict(),
'optimizer': optimizer.state_dict()
}, weight_with_optimizer)
torch.save({'net': net.state_dict()}, weight)
print('saved')
def get_dataloaders(data_location, labels_file, modalities,
wsi_patch_size=None, n_wsi_patches=None, batch_size=None,
exclude_patients=None, return_patient_id=False):
"""Instantiate PyTorch DataLoaders.
Parameters
----------
Returns
-------
Dict of Pytorch Dataloaders.
"""
data_dirs = {
'clinical': os.path.join(data_location, 'Clinical'),
'wsi': os.path.join(data_location, 'WSI'),
'mRNA': os.path.join(data_location, 'RNA-seq'),
'miRNA': os.path.join(data_location, 'miRNA-seq'),
'DNAm': os.path.join(data_location, 'DNAm/5k'),
'CNV': os.path.join(data_location, 'CNV'),
}
data_dirs = {mod: data_dirs[mod] for mod in modalities}
if batch_size is None:
if 'wsi' in data_dirs.keys() and n_wsi_patches > 1:
batch_size = 2**5
else:
batch_size = 2**7
patient_labels = {'train': get_label_map(labels_file, 'train'),
'val': get_label_map(labels_file, 'val'),
'test': get_label_map(labels_file, 'test')}
if 'wsi' in data_dirs.keys():
transforms = {
'train': torchvision.transforms.Compose([
patch_transforms.ToPIL(),
torchvision.transforms.CenterCrop(wsi_patch_size),
torchvision.transforms.ColorJitter(
brightness=64/255, contrast=0.5, saturation=0.25,
hue=0.04),
patch_transforms.ToNumpy(),
patch_transforms.RandomRotate(),
patch_transforms.RandomFlipUpDown(),
patch_transforms.ToTensor(),
]),
# No data augmentation for validation
'val': torchvision.transforms.Compose([
patch_transforms.ToPIL(),
torchvision.transforms.CenterCrop(wsi_patch_size),
patch_transforms.ToNumpy(),
patch_transforms.ToTensor(),
]),
'test': torchvision.transforms.Compose([
patch_transforms.ToPIL(),
torchvision.transforms.CenterCrop(wsi_patch_size),
patch_transforms.ToNumpy(),
patch_transforms.ToTensor(),
])}
else:
transforms = {'train': None, 'val': None, 'test': None}
datasets = {x: dataset.MultimodalDataset(
label_map=patient_labels[x],
data_dirs=data_dirs,
n_patches=n_wsi_patches,
patch_size=wsi_patch_size,
transform=transforms[x],
exclude_patients=exclude_patients,
return_patient_id=return_patient_id)
for x in ['train', 'val', 'test']}
print('Data modalities:')
for mod in modalities:
print(' ', mod)
print()
print('Dataset sizes (# patients):')
for x in datasets.keys():
print(f' {x}: {len(datasets[x])}')
print()
print('Batch size:', batch_size)
# Use "drop_last=True" to drop the last incomplete batch
# to avoid undefined loss values due to lack of sufficient
# orderable observation pairs caused by data censorship
# When running all data with batch = 64:
# 8880 % 64 = 48
# When running 20 cancer data with batch = 64:
# 7369 % 64 = 9
dataloaders = {'train': torch.utils.data.DataLoader(
datasets['train'], batch_size=batch_size,
shuffle=True, num_workers=4, drop_last=True),
'val': torch.utils.data.DataLoader(
datasets['val'], batch_size=batch_size * 2,
shuffle=False, num_workers=4, drop_last=True),
'test': torch.utils.data.DataLoader(
datasets['test'], batch_size=batch_size * 2,
shuffle=False, num_workers=4, drop_last=True)}
return dataloaders
def train():
transforms = [
Transforms.RondomFlip(),
Transforms.RandomRotate(15),
Transforms.Log(0.5),
Transforms.Blur(0.2),
Transforms.ToGray(),
Transforms.ToTensor()
]
train_dataset = UNetDataset('./data/train/',
'./data/train_cleaned/',
transform=transforms)
train_dataLoader = DataLoader(dataset=train_dataset,
batch_size=config.BATCH_SIZE,
shuffle=True,
num_workers=0)
valid_dataset = UNetDataset('./data/valid/',
'./data/valid_cleaned/',
transform=transforms)
valid_dataLoader = DataLoader(dataset=valid_dataset,
batch_size=config.BATCH_SIZE,
shuffle=True,
num_workers=0)
net = UNet(n_channels=config.n_channels,
n_classes=config.n_classes).to(config.device)
writer = SummaryWriter()
optimizer = optim.Adam(net.parameters(), lr=config.LR)
if config.n_classes > 1:
loss_func = nn.CrossEntropyLoss().to(config.device)
else:
loss_func = nn.BCEWithLogitsLoss().to(config.device)
best_loss = float('inf')
if os.path.exists(config.weight_with_optimizer):
checkpoint = torch.load(config.weight_with_optimizer,
map_location='cpu')
net.load_state_dict(checkpoint['net'])
optimizer.load_state_dict(checkpoint['optimizer'])
print('load weight')
for epoch in range(config.EPOCH):
train_loss = 0
net.train()
for step, (batch_x, batch_y) in enumerate(train_dataLoader):
batch_x = batch_x.to(device=config.device)
batch_y = batch_y.squeeze(1).to(device=config.device)
output = net(batch_x)
loss = loss_func(output, batch_y)
train_loss += loss.item()
if loss < best_loss:
best_loss = loss
torch.save(
{
'net': net.state_dict(),
'optimizer': optimizer.state_dict()
}, config.best_model_with_optimizer)
torch.save({'net': net.state_dict()}, config.best_model)
optimizer.zero_grad()
loss.backward()
optimizer.step()
net.eval()
eval_loss = 0
for step, (batch_x, batch_y) in enumerate(valid_dataLoader):
batch_x = batch_x.to(device=config.device)
batch_y = batch_y.squeeze(1).to(device=config.device)
output = net(batch_x)
valid_loss = loss_func(output, batch_y)
eval_loss += valid_loss.item()
writer.add_scalar("train_loss", train_loss, epoch)
writer.add_scalar("eval_loss", eval_loss, epoch)
print("*" * 80)
print('epoch: %d | train loss: %.4f | valid loss: %.4f' %
(epoch, train_loss, eval_loss))
print("*" * 80)
if (epoch + 1) % 10 == 0:
torch.save(
{
'net': net.state_dict(),
'optimizer': optimizer.state_dict()
}, config.weight_with_optimizer)
torch.save({'net': net.state_dict()}, config.weight)
print('saved')
writer.close()
weights[foreground] = n * self.fg_weight / foreground.sum().item()
mask = mask * weights
img = torch.cat([img, self.mean[None]], dim=0)
if self.flip:
img, mask, lbl = random_flip(img, mask, lbl)
return img, mask, lbl
ROTATE_TRANS_1024 = tr.Compose([
tr.AspectPreservingResizeTransform((1024, 768)),
tr.Lift(T.Pad(88)),
tr.RandomRotate(),
])
PAD_TRANS_1024 = tr.Compose([
tr.AspectPreservingResizeTransform((1024, 768)),
tr.Lift(T.Pad(88)),
])
RotatedISICDataset = rotated_dataset(ISICDataset)
if __name__ == '__main__':
target_size = 1024, 768
img_transform = T.Compose([T.ColorJitter(0.3, 0.3, 0.3, 0.), T.ToTensor()])
d = ISICDataset('/home/jatentaki/Storage/jatentaki/Datasets/isic2018',
'print_aux': True,
}
ckpt_path = './ckpt'
exp_name = 'model'
img_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize([0.3598, 0.3653, 0.3662], [0.2573, 0.2663, 0.2756])
])
mask_transform = extend_transforms.MaskToTensor()
train_joint_transform = extend_transforms.Compose([
extend_transforms.RandomScale(),
extend_transforms.RandomSizedRatio(760, 842, 274, 304),
extend_transforms.RandomRotate(args['rotate_degree']),
extend_transforms.RandomCrop(args['train_crop_size']),
])
train_set = culane.CULANE('train',
joint_transform=train_joint_transform,
transform=img_transform,
mask_transform=mask_transform)
train_loader = DataLoader(train_set,
batch_size=args['train_batch_size'],
num_workers=10,
shuffle=True)
criterion = CrossEntropyLoss2d(weight=torch.Tensor([0.4, 1, 1, 1, 1]).cuda(),
size_average=True,
ignore_index=culane.ignore_label,
print("Ignoring --epochs outside of training mode")
if args.no_jit and args.optimize:
print("Ignoring --optimize in --no-jit setting")
writer.add_text('general', str(vars(args)))
transform = T.Compose([T.CenterCrop(644), T.ToTensor()])
# if we are not padding the convolutions, we have to pad the input
aug_pad = None if args.padding else tr.Lift(T.Pad(40))
test_global_transform = aug_pad
tr_global_transform = tr.Compose(
[tr.RandomRotate(), tr.RandomFlip(), aug_pad])
train_data = loader.DriveDataset(args.data_path,
training=True,
bloat=args.bloat,
from_=args.cut,
img_transform=transform,
mask_transform=transform,
label_transform=transform,
global_transform=tr_global_transform)
train_loader = DataLoader(train_data,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.workers)
if args.test_on_train: