def runInference(args: argparse.Namespace):
print('>>> Loading model')
net = torch.load(args.model_weights)
device = torch.device("cuda")
net.to(device)
print('>>> Loading the data')
batch_size: int = args.batch_size
num_classes: int = args.num_classes
transform = transforms.Compose([
lambda img: np.array(img)[np.newaxis, ...],
lambda nd: nd / 255, # max <= 1
lambda nd: torch.tensor(nd, dtype=torch.float32)
])
folders: List[Path] = [Path(args.data_folder)]
names: List[str] = map_(lambda p: str(p.name), folders[0].glob("*.png"))
dt_set = SliceDataset(names,
folders,
transforms=[transform],
debug=False,
C=num_classes)
loader = DataLoader(dt_set,
batch_size=batch_size,
num_workers=batch_size + 2,
shuffle=False,
drop_last=False)
print('>>> Starting the inference')
savedir: str = args.save_folder
total_iteration = len(loader)
desc = f">> Inference"
tq_iter = tqdm_(enumerate(loader), total=total_iteration, desc=desc)
with torch.no_grad():
for j, (filenames, image, _) in tq_iter:
image = image.to(device)
pred_logits: Tensor = net(image)
pred_probs: Tensor = F.softmax(pred_logits, dim=1)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
predicted_class: Tensor = probs2class(pred_probs)
save_images(predicted_class, filenames, savedir, "", 0)
def runInference(args: argparse.Namespace):
print('>>> Loading model')
net = torch.load(args.model_weights)
device = torch.device("cuda")
net.to(device)
print('>>> Loading the data')
batch_size: int = args.batch_size
num_classes: int = args.num_classes
folders: list[Path] = [Path(args.data_folder)]
names: list[str] = map_(lambda p: str(p.name), folders[0].glob("*.png"))
dt_set = SliceDataset(
names,
folders * 2, # Duplicate for compatibility reasons
are_hots=[False, False],
transforms=[png_transform, dummy_gt_transform
], # So it is happy about the target size
bounds_generators=[],
debug=args.debug,
K=num_classes)
loader = DataLoader(dt_set,
batch_size=batch_size,
num_workers=batch_size + 2,
shuffle=False,
drop_last=False,
collate_fn=custom_collate)
print('>>> Starting the inference')
savedir: Path = Path(args.save_folder)
savedir.mkdir(parents=True, exist_ok=True)
total_iteration = len(loader)
desc = ">> Inference"
tq_iter = tqdm_(enumerate(loader), total=total_iteration, desc=desc)
with torch.no_grad():
for j, data in tq_iter:
filenames: list[str] = data["filenames"]
image: Tensor = data["images"].to(device)
pred_logits: Tensor = net(image)
pred_probs: Tensor = F.softmax(pred_logits, dim=1)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
predicted_class: Tensor
if args.mode == 'argmax':
predicted_class = probs2class(pred_probs)
elif args.mode == 'probs':
predicted_class = (pred_probs[:, args.probs_class, ...] *
255).type(torch.uint8)
elif args.mode == 'threshold':
thresholded: Tensor = pred_probs[:, ...] > args.threshold
predicted_class = thresholded.argmax(dim=1)
elif args.mode == 'softmax':
for i, filename in enumerate(filenames):
np.save((savedir / filename).with_suffix(".npy"),
pred_probs[i].cpu().numpy())
if args.mode != 'softmax':
save_images(predicted_class, filenames, savedir)
def do_epoch(args, mode: str, net: Any, device: Any, loader: DataLoader, epc: int,
loss_fns: List[Callable], loss_weights: List[float],loss_fns_source: List[Callable],
loss_weights_source: List[float], new_w:int, num_steps:int, C: int, metric_axis:List[int], savedir: str = "",
optimizer: Any = None, target_loader: Any = None):
assert mode in ["train", "val"]
L: int = len(loss_fns)
indices = torch.tensor(metric_axis,device=device)
if mode == "train":
net.train()
desc = f">> Training ({epc})"
elif mode == "val":
net.eval()
# net.train()
desc = f">> Validation ({epc})"
total_it_s, total_images = len(loader), len(loader.dataset)
total_it_t, total_images_t = len(target_loader), len(target_loader.dataset)
total_iteration = max(total_it_s, total_it_t)
# Lazy add lines below because we will be cycling until the biggest length is reached
total_images = max(total_images, total_images_t)
total_images_t = total_images
pho=1
dtype = eval(args.dtype)
all_dices: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_inter_card: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_card_gt: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_card_pred: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
loss_log: Tensor = torch.zeros((total_images), dtype=dtype, device=device)
loss_inf: Tensor = torch.zeros((total_images), dtype=dtype, device=device)
loss_cons: Tensor = torch.zeros((total_images), dtype=dtype, device=device)
loss_fs: Tensor = torch.zeros((total_images), dtype=dtype, device=device)
posim_log: Tensor = torch.zeros((total_images), dtype=dtype, device=device)
haussdorf_log: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_grp: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
dice_3d_log: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
dice_3d_sd_log: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
if args.source_metrics == True:
all_dices_s: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_inter_card_s: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_card_gt_s: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_card_pred_s: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_grp_s: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
dice_3d_s_log: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
dice_3d_s_sd_log: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
# if len(loader)>len(target_loader):
# tq_iter = tqdm_(enumerate(zip(loader, cycle(target_loader))), total=total_iteration, desc=desc)
# elif len(loader)<len(target_loader):
# tq_iter = tqdm_(enumerate(zip(cycle(loader), target_loader)), total=total_iteration, desc=desc)
# else:
# tq_iter = tqdm_(enumerate(zip(loader, target_loader)), total=total_iteration, desc=desc)
tq_iter = tqdm_(enumerate(zip(loader, target_loader)), total=total_iteration, desc=desc)
#tq_iter = tqdm_(enumerate(target_loader), total=total_iteration, desc=desc)
done: int = 0
#ratio_losses = 0
n_warmup = 0
mult_lw = [pho ** (epc - n_warmup + 1)] * len(loss_weights)
#if epc > 100:
# mult_lw = [pho ** 100] * len(loss_weights)
mult_lw[0] = 1
loss_weights = [a * b for a, b in zip(loss_weights, mult_lw)]
losses_vec, source_vec, target_vec, baseline_target_vec = [], [], [], []
pen_count = 0
with warnings.catch_warnings():
warnings.simplefilter("ignore")
for j, (source_data, target_data) in tq_iter:
#for j, target_data in tq_iter:
source_data[1:] = [e.to(device) for e in source_data[1:]] # Move all tensors to device
filenames_source, source_image, source_gt = source_data[:3]
target_data[1:] = [e.to(device) for e in target_data[1:]] # Move all tensors to device
filenames_target, target_image, target_gt = target_data[:3]
labels = target_data[3:3+L]
labels_source = source_data[3:3 + L]
bounds = target_data[3+L:]
bounds_source = source_data[3+L:]
assert len(labels) == len(bounds), len(bounds)
if args.mix == False:
assert filenames_source == filenames_target
#print(filenames_source,filenames_target)
B = len(target_image)
# Reset gradients
if optimizer:
#adjust_learning_rate(optimizer, 1, args.l_rate, num_steps, args.power)
optimizer.zero_grad()
# Forward
with torch.set_grad_enabled(mode == "train"):
pred_logits: Tensor = net(target_image)
pred_logits_source: Tensor = net(source_image)
pred_probs: Tensor = F.softmax(pred_logits, dim=1)
pred_probs_source: Tensor = F.softmax(pred_logits_source, dim=1)
if new_w > 0:
pred_probs = resize(pred_probs, new_w)
labels = [resize(label, new_w) for label in labels]
target = resize(target, new_w)
predicted_mask: Tensor = probs2one_hot(pred_probs) # Used only for dice computation
predicted_mask_source: Tensor = probs2one_hot(pred_probs_source) # Used only for dice computation
#print(torch.sum(predicted_mask, dim=[2,3]).cpu().numpy())
#print(list(map(lambda n: [int(f) for f in n], np.around(torch.sum(pred_probs, dim=[2,3]).detach().cpu().numpy()))))
assert len(bounds) == len(loss_fns) == len(loss_weights)
if epc < n_warmup:
loss_weights = [0]*len(loss_weights)
loss: Tensor = torch.zeros(1, requires_grad=True).to(device)
loss_vec = []
loss_k = []
for loss_fn,label, w, bound in zip(loss_fns,labels, loss_weights, bounds):
if w > 0:
if args.lin_aug_w:
if epc<70:
w=w*(epc+1)/70
loss_b = loss_fn(pred_probs, label, bound)
loss = loss + w * loss_b
#pen_count += count_b.detach()
#print(count_b.detach())
loss_k.append(w*loss_b.detach())
#for loss_fn, label, w, bound in zip(loss_fns_source, [source_gt], loss_weights_source, torch.randn(1)):
#for loss_fn, label, w, bound in zip(loss_fns_source, labels_source, loss_weights_source, torch.randn(1)):
for loss_fn, label, w, bound in zip(loss_fns_source, labels_source, loss_weights_source, bounds_source):
if w > 0:
loss_b = loss_fn(pred_probs_source, label, bound)
loss = loss+ w * loss_b
loss_k.append(w*loss_b.detach())
#print(loss_k)
# Backward
if optimizer:
loss.backward()
optimizer.step()
# Compute and log metrics
#dices: Tensor = dice_coef(predicted_mask.detach(), target.detach())
# baseline_dices: Tensor = dice_coef(labels[0].detach(), target.detach())
#batch_dice: Tensor = dice_batch(predicted_mask.detach(), target.detach())
# assert batch_dice.shape == (C,) and dices.shape == (B, C), (batch_dice.shape, dices.shape, B, C)
dices, inter_card, card_gt, card_pred = dice_coef(predicted_mask.detach(), target_gt.detach())
assert dices.shape == (B, C), (dices.shape, B, C)
sm_slice = slice(done, done + B) # Values only for current batch
all_dices[sm_slice, ...] = dices
# # for 3D dice
all_grp[sm_slice, ...] = int(re.split('_', filenames_target[0])[1]) * torch.ones([1, C])
all_inter_card[sm_slice, ...] = inter_card
all_card_gt[sm_slice, ...] = card_gt
all_card_pred[sm_slice, ...] = card_pred
# 3D dice on source
if args.source_metrics ==True:
dices_s, inter_card_s, card_gt_s, card_pred_s = dice_coef(predicted_mask_source.detach(), source_gt.detach())
all_grp_s[sm_slice, ...] = int(re.split('_', filenames_source[0])[1]) * torch.ones([1, C])
all_inter_card_s[sm_slice, ...] = inter_card_s
all_card_gt_s[sm_slice, ...] = card_gt_s
all_card_pred_s[sm_slice, ...] = card_pred_s
#loss_log[sm_slice] = loss.detach()
loss_inf[sm_slice] = loss_k[0]
if len(loss_k)>1:
loss_cons[sm_slice] = loss_k[1]
else:
loss_cons[sm_slice] = 0
if len(loss_k)>2:
loss_fs[sm_slice] = loss_k[2]
else:
loss_fs[sm_slice] = 0
#posim_log[sm_slice] = torch.einsum("bcwh->b", [target_gt[:, 1:, :, :]]).detach() > 0
#haussdorf_res: Tensor = haussdorf(predicted_mask.detach(), target_gt.detach(), dtype)
#assert haussdorf_res.shape == (B, C)
#haussdorf_log[sm_slice] = haussdorf_res
# # Save images
if savedir:
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=UserWarning)
warnings.simplefilter("ignore")
predicted_class: Tensor = probs2class(pred_probs)
#save_images_p(predicted_class, filenames_target, args.dataset, mode, epc, False)
save_images(predicted_class, filenames_target, savedir, mode, epc, True)
# Logging
big_slice = slice(0, done + B) # Value for current and previous batches
stat_dict = {"dice": torch.index_select(all_dices, 1, indices).mean(),
"loss": loss_log[big_slice].mean()}
nice_dict = {k: f"{v:.4f}" for (k, v) in stat_dict.items()}
done += B
tq_iter.set_postfix(nice_dict)
print(f"{desc} " + ', '.join(f"{k}={v}" for (k, v) in nice_dict.items()))
#dice_posim = torch.masked_select(all_dices[:, -1], posim_log.type(dtype=torch.uint8)).mean()
# dice3D gives back the 3d dice mai on images
# if not args.debug:
# dice_3d_log_o, dice_3d_sd_log_o = dice3d(args.workdir, f"iter{epc:03d}", mode, "Subj_\\d+_",args.dataset + mode + '/CT_GT', C)
dice_3d_log, dice_3d_sd_log = dice3dn(all_grp, all_inter_card, all_card_gt, all_card_pred,metric_axis,True)
if args.source_metrics ==True:
dice_3d_s_log, dice_3d_s_sd_log = dice3dn(all_grp_s, all_inter_card_s, all_card_gt_s, all_card_pred_s,metric_axis,True)
print("mean 3d_dice over all patients:",dice_3d_log)
#source_vec = [ dice_3d_s, dice_3d_sd_s, haussdorf_log_s]
dice_2d = torch.index_select(all_dices, 1, indices).mean().cpu().numpy()
target_vec = [ dice_3d_log, dice_3d_sd_log, dice_2d]
if args.source_metrics ==True:
source_vec = [ dice_3d_s_log, dice_3d_s_sd_log]
else:
source_vec = [0,0]
#losses_vec = [loss_log.mean().item()]
losses_vec = [loss_inf.mean().item(),loss_cons.mean().item(),loss_fs.mean().item()]
return losses_vec, target_vec,source_vec
theta, forward = LeNet5(
batch_size=batch_size,
num_particles=num_particles,
)
@jit
def loss(theta, x, y):
yhat = forward(theta, x)
return cross_entropy(y, yhat)
optimizer = SGD(0.001)
grad_loss = jit(grad(loss))
# SVGD training
for epoch in range(100):
for i, (x, y) in tqdm_(train_loader):
g = get_phi(theta, grad_loss(theta, x.numpy(), y.numpy()))
theta = optimizer.update(theta, g)
test_acc = []
for i, (x, y) in tqdm_(test_loader):
yhat = forward(theta, x.numpy())
nll = cross_entropy(y.numpy(), yhat)
pred = yhat.mean(axis=0)
correct = (pred.argmax(axis=1) == y.numpy()).mean()
test_acc.append(float(correct))
print("Iteration: ", epoch, "Cross Entropy:", nll, "Test Accuracy:", np_normal.mean(test_acc))
def runInference(args: argparse.Namespace, pred_folder: str):
# print('>>> Loading the data')
device = torch.device("cuda") if torch.cuda.is_available(
) and not args.cpu else torch.device("cpu")
C: int = args.num_classes
# Let's just reuse some code
png_transform = transforms.Compose([
lambda img: np.array(img)[np.newaxis, ...],
lambda nd: nd / 255, # max <= 1
lambda nd: torch.tensor(nd, dtype=torch.float32)
])
gt_transform = transforms.Compose([
lambda img: np.array(img)[np.newaxis, ...],
lambda nd: torch.tensor(nd, dtype=torch.int64),
partial(class2one_hot, C=C),
itemgetter(0)
])
bounds_gen = [(lambda *a: torch.zeros(C, 1, 2)) for _ in range(2)]
folders: List[Path] = [
Path(pred_folder),
Path(pred_folder),
Path(args.gt_folder)
] # First one is dummy
names: List[str] = map_(lambda p: str(p.name), folders[0].glob("*.png"))
are_hots = [False, True, True]
dt_set = SliceDataset(
names,
folders,
transforms=[png_transform, gt_transform, gt_transform],
debug=False,
C=C,
are_hots=are_hots,
in_memory=False,
bounds_generators=bounds_gen)
sampler = PatientSampler(dt_set, args.grp_regex)
loader = DataLoader(dt_set, batch_sampler=sampler, num_workers=11)
# print('>>> Computing the metrics')
total_iteration, total_images = len(loader), len(loader.dataset)
metrics = {
"all_dices":
torch.zeros((total_images, C), dtype=torch.float64, device=device),
"batch_dices":
torch.zeros((total_iteration, C), dtype=torch.float64, device=device),
"sizes":
torch.zeros((total_images, 1), dtype=torch.float64, device=device)
}
desc = f">> Computing"
tq_iter = tqdm_(enumerate(loader), total=total_iteration, desc=desc)
done: int = 0
for j, (filenames, _, pred, gt, _) in tq_iter:
B = len(pred)
pred = pred.to(device)
gt = gt.to(device)
assert simplex(pred) and sset(pred, [0, 1])
assert simplex(gt) and sset(gt, [0, 1])
dices: Tensor = dice_coef(pred, gt)
b_dices: Tensor = dice_batch(pred, gt)
assert dices.shape == (B, C)
assert b_dices.shape == (C, ), b_dices.shape
sm_slice = slice(done, done + B) # Values only for current batch
metrics["all_dices"][sm_slice, ...] = dices
metrics["sizes"][sm_slice, :] = torch.einsum("bwh->b",
gt[:, 1, ...])[..., None]
metrics["batch_dices"][j] = b_dices
done += B
print(f">>> {pred_folder}")
for key, v in metrics.items():
print(key, map_("{:.4f}".format, v.mean(dim=0)))
def do_epoch(mode: str, net: Any, device: Any, loader: DataLoader, epc: int,
loss_fns: List[Callable], loss_weights: List[float], C: int,
savedir: str = "", optimizer: Any = None,
metric_axis: List[int] = [1], compute_haussdorf: bool = False) \
-> Tuple[Tensor, Tensor, Tensor, Tensor]:
assert mode in ["train", "val"]
L: int = len(loss_fns)
if mode == "train":
net.train()
desc = f">> Training ({epc})"
elif mode == "val":
net.eval()
desc = f">> Validation ({epc})"
total_iteration, total_images = len(loader), len(loader.dataset)
all_dices: Tensor = torch.zeros((total_images, C),
dtype=torch.float32,
device=device)
batch_dices: Tensor = torch.zeros((total_iteration, C),
dtype=torch.float32,
device=device)
loss_log: Tensor = torch.zeros((total_iteration),
dtype=torch.float32,
device=device)
haussdorf_log: Tensor = torch.zeros((total_images, C),
dtype=torch.float32,
device=device)
tq_iter = tqdm_(enumerate(loader), total=total_iteration, desc=desc)
done: int = 0
for j, data in tq_iter:
data[1:] = [e.to(device)
for e in data[1:]] # Move all tensors to device
filenames, image, target = data[:3]
labels = data[3:3 + L]
bounds = data[3 + L:]
assert len(labels) == len(bounds)
B = len(image)
# Reset gradients
if optimizer:
optimizer.zero_grad()
# Forward
pred_logits: Tensor = net(image)
pred_probs: Tensor = F.softmax(pred_logits, dim=1)
predicted_mask: Tensor = probs2one_hot(
pred_probs.detach()) # Used only for dice computation
assert len(bounds) == len(loss_fns) == len(loss_weights)
ziped = zip(loss_fns, labels, loss_weights, bounds)
losses = [
w * loss_fn(pred_probs, label, bound)
for loss_fn, label, w, bound in ziped
]
loss = reduce(add, losses)
assert loss.shape == (), loss.shape
# Backward
if optimizer:
loss.backward()
optimizer.step()
# Compute and log metrics
loss_log[j] = loss.detach()
sm_slice = slice(done, done + B) # Values only for current batch
dices: Tensor = dice_coef(predicted_mask, target.detach())
assert dices.shape == (B, C), (dices.shape, B, C)
all_dices[sm_slice, ...] = dices
if B > 1 and mode == "val":
batch_dice: Tensor = dice_batch(predicted_mask, target.detach())
assert batch_dice.shape == (C, ), (batch_dice.shape, B, C)
batch_dices[j] = batch_dice
if compute_haussdorf:
haussdorf_res: Tensor = haussdorf(predicted_mask.detach(),
target.detach())
assert haussdorf_res.shape == (B, C)
haussdorf_log[sm_slice] = haussdorf_res
# Save images
if savedir:
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=UserWarning)
predicted_class: Tensor = probs2class(pred_probs)
save_images(predicted_class, filenames, savedir, mode, epc)
# Logging
big_slice = slice(0,
done + B) # Value for current and previous batches
dsc_dict = {
f"DSC{n}": all_dices[big_slice, n].mean()
for n in metric_axis
}
hauss_dict = {
f"HD{n}": haussdorf_log[big_slice, n].mean()
for n in metric_axis
} if compute_haussdorf else {}
batch_dict = {
f"bDSC{n}": batch_dices[:j, n].mean()
for n in metric_axis
} if B > 1 and mode == "val" else {}
mean_dict = {
"DSC": all_dices[big_slice, metric_axis].mean(),
"HD": haussdorf_log[big_slice, metric_axis].mean()
} if len(metric_axis) > 1 else {}
stat_dict = {
**dsc_dict,
**hauss_dict,
**mean_dict,
**batch_dict, "loss": loss_log[:j].mean()
}
nice_dict = {k: f"{v:.3f}" for (k, v) in stat_dict.items()}
tq_iter.set_postfix(nice_dict)
done += B
print(f"{desc} " + ', '.join(f"{k}={v}" for (k, v) in nice_dict.items()))
return loss_log, all_dices, batch_dices, haussdorf_log
def do_epoch(mode: str, net: Any, device: Any, loaders: List[DataLoader], epc: int,
list_loss_fns: List[List[Callable]], list_loss_weights: List[List[float]], C: int,
savedir: str = "", optimizer: Any = None,
metric_axis: List[int] = [1], compute_haussdorf: bool = False, compute_miou: bool = False,
temperature: float = 1) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tuple[None, Tensor]]:
assert mode in ["train", "val"]
if mode == "train":
net.train()
desc = f">> Training ({epc})"
elif mode == "val":
net.eval()
desc = f">> Validation ({epc})"
total_iteration: int = sum(len(loader) for loader in loaders) # U
total_images: int = sum(len(loader.dataset) for loader in loaders) # D
n_loss: int = max(map(len, list_loss_fns))
all_dices: Tensor = torch.zeros((total_images, C), dtype=torch.float32, device=device)
batch_dices: Tensor = torch.zeros((total_iteration, C), dtype=torch.float32, device=device)
loss_log: Tensor = torch.zeros((total_iteration, n_loss), dtype=torch.float32, device=device)
haussdorf_log: Tensor = torch.zeros((total_images, C), dtype=torch.float32, device=device)
iiou_log: Tensor = torch.zeros((total_images, C), dtype=torch.float32, device=device)
intersections: Tensor = torch.zeros((total_images, C), dtype=torch.float32, device=device)
unions: Tensor = torch.zeros((total_images, C), dtype=torch.float32, device=device)
few_axis: bool = len(metric_axis) <= 3
done_img: int = 0
done_batch: int = 0
tq_iter = tqdm_(total=total_iteration, desc=desc)
for i, (loader, loss_fns, loss_weights) in enumerate(zip(loaders, list_loss_fns, list_loss_weights)):
L: int = len(loss_fns)
for data in loader:
data[1:] = [e.to(device) for e in data[1:]] # Move all tensors to device
filenames, image, target = data[:3]
assert not target.requires_grad
labels = data[3:3 + L]
bounds = data[3 + L:]
assert len(labels) == len(bounds)
B = len(image)
# Reset gradients
if optimizer:
optimizer.zero_grad()
# Forward
pred_logits: Tensor = net(image)
pred_probs: Tensor = F.softmax(temperature * pred_logits, dim=1)
predicted_mask: Tensor = probs2one_hot(pred_probs.detach()) # Used only for dice computation
assert not predicted_mask.requires_grad
assert len(bounds) == len(loss_fns) == len(loss_weights) == len(labels)
ziped = zip(loss_fns, labels, loss_weights, bounds)
losses = [w * loss_fn(pred_probs, label, bound) for loss_fn, label, w, bound in ziped]
loss = reduce(add, losses)
assert loss.shape == (), loss.shape
# if epc >= 1 and False:
# import matplotlib.pyplot as plt
# _, axes = plt.subplots(nrows=1, ncols=3)
# axes[0].imshow(image[0, 0].cpu().numpy(), cmap='gray')
# axes[0].contour(target[0, 1].cpu().numpy(), cmap='rainbow')
# pred_np = pred_probs[0, 1].detach().cpu().numpy()
# axes[1].imshow(pred_np)
# bins = np.linspace(0, 1, 50)
# axes[2].hist(pred_np.flatten(), bins)
# print(bounds)
# print(bounds[2].cpu().numpy())
# print(bounds[2][0, 1].cpu().numpy())
# print(pred_np.sum())
# plt.show()
# Backward
if optimizer:
loss.backward()
optimizer.step()
# Compute and log metrics
# loss_log[done_batch] = loss.detach()
for j in range(len(loss_fns)):
loss_log[done_batch, j] = losses[j].detach()
sm_slice = slice(done_img, done_img + B) # Values only for current batch
dices: Tensor = dice_coef(predicted_mask, target)
assert dices.shape == (B, C), (dices.shape, B, C)
all_dices[sm_slice, ...] = dices
if B > 1 and mode == "val":
batch_dice: Tensor = dice_batch(predicted_mask, target)
assert batch_dice.shape == (C,), (batch_dice.shape, B, C)
batch_dices[done_batch] = batch_dice
if compute_haussdorf:
haussdorf_res: Tensor = haussdorf(predicted_mask, target)
assert haussdorf_res.shape == (B, C)
haussdorf_log[sm_slice] = haussdorf_res
if compute_miou:
IoUs: Tensor = iIoU(predicted_mask, target)
assert IoUs.shape == (B, C), IoUs.shape
iiou_log[sm_slice] = IoUs
intersections[sm_slice] = inter_sum(predicted_mask, target)
unions[sm_slice] = union_sum(predicted_mask, target)
# Save images
if savedir:
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=UserWarning)
predicted_class: Tensor = probs2class(pred_probs)
save_images(predicted_class, filenames, savedir, mode, epc)
# Logging
big_slice = slice(0, done_img + B) # Value for current and previous batches
dsc_dict = {f"DSC{n}": all_dices[big_slice, n].mean() for n in metric_axis} if few_axis else {}
hauss_dict = {f"HD{n}": haussdorf_log[big_slice, n].mean() for n in metric_axis} \
if compute_haussdorf and few_axis else {}
batch_dict = {f"bDSC{n}": batch_dices[:done_batch, n].mean() for n in metric_axis} \
if B > 1 and mode == "val" and few_axis else {}
miou_dict = {f"iIoU": iiou_log[big_slice, metric_axis].mean(),
f"mIoU": (intersections.sum(dim=0) / (unions.sum(dim=0) + 1e-10)).mean()} \
if compute_miou else {}
if len(metric_axis) > 1:
mean_dict = {"DSC": all_dices[big_slice, metric_axis].mean()}
if compute_haussdorf:
mean_dict["HD"] = haussdorf_log[big_slice, metric_axis].mean()
else:
mean_dict = {}
stat_dict = {**miou_dict, **dsc_dict, **hauss_dict, **mean_dict, **batch_dict,
"loss": loss_log[:done_batch].mean()}
nice_dict = {k: f"{v:.3f}" for (k, v) in stat_dict.items()}
done_img += B
done_batch += 1
tq_iter.set_postfix({**nice_dict, "loader": str(i)})
tq_iter.update(1)
tq_iter.close()
print(f"{desc} " + ', '.join(f"{k}={v}" for (k, v) in nice_dict.items()))
if compute_miou:
mIoUs: Tensor = (intersections.sum(dim=0) / (unions.sum(dim=0) + 1e-10))
assert mIoUs.shape == (C,), mIoUs.shape
else:
mIoUs = None
if not few_axis and False:
print(f"DSC: {[f'{all_dices[:, n].mean():.3f}' for n in metric_axis]}")
print(f"iIoU: {[f'{iiou_log[:, n].mean():.3f}' for n in metric_axis]}")
if mIoUs:
print(f"mIoU: {[f'{mIoUs[n]:.3f}' for n in metric_axis]}")
return loss_log, all_dices, batch_dices, haussdorf_log, mIoUs
def do_epc(epc: int, mode: str, net: Any, loader: DataLoader, device, criterion, args,
optimizer: Any = None) -> Tuple[Any, Dict[str, Tensor]]:
assert mode in ["train", "val"]
desc: str
if mode == "train":
net.train()
desc = f">> Training ({epc})"
elif mode == "val":
net.eval()
desc = f">> Validation ({epc})"
total_iteration: int = len(loader) # U
total_images: int = len(loader.dataset) # D
metrics = {"loss": torch.zeros((total_iteration), dtype=torch.float32, device=device),
"diff": torch.zeros((total_images, args.n_class), dtype=torch.float32, device=device)}
tq_iter = tqdm_(total=total_iteration, desc=desc)
done_img: int = 0
for j, data in enumerate(loader):
data[1:] = [e.to(device) for e in data[1:]] # Move all tensors to device
# filenames, images, targets = data[:3]
filenames, images, targets = data
assert len(filenames) == len(images) == len(targets)
B: int = len(images)
sizes = einsum("bcwh->bc", targets).type(torch.float32)
if optimizer:
optimizer.zero_grad()
if args.pretrained:
b, c, w, h = images.shape
assert c == 1
viewed = images.view((b, w, h))
new_img = torch.stack([viewed, viewed, viewed], dim=1)
assert new_img.shape == (b, 3, w, h), new_img.shape
images = new_img
predicted_sizes = net(images)
assert sizes.shape == predicted_sizes.shape
loss = criterion(predicted_sizes[:, args.idc], sizes[:, args.idc])
if optimizer:
loss.backward()
optimizer.step()
metrics["loss"][j] = loss.detach().item()
metrics["diff"][done_img:done_img + B, ...] = torch.abs(predicted_sizes.detach() - sizes.detach())[...]
stat_dict: Dict = {"loss": metrics["loss"][:j].mean(),
"diff": metrics["diff"][:done_img + B, args.idc].mean()}
nice_dict: Dict = {k: f"{v:12.2f}" for (k, v) in stat_dict.items()}
done_img += B
tq_iter.set_postfix(nice_dict)
tq_iter.update(1)
tq_iter.close()
print(f"{desc} " + ', '.join(f"{k}={v}" for (k, v) in nice_dict.items()))
return net, metrics
def do_epoch(
mode: str,
net: Any,
device: Any,
loaders: list[DataLoader],
epc: int,
list_loss_fns: list[list[Callable]],
list_loss_weights: list[list[float]],
K: int,
savedir: str = "",
optimizer: Any = None,
metric_axis: list[int] = [1],
compute_3d_dice: bool = False,
temperature: float = 1) -> Tuple[Tensor, Tensor, Optional[Tensor]]:
assert mode in ["train", "val", "dual"]
if mode == "train":
net.train()
desc = f">> Training ({epc})"
elif mode == "val":
net.eval()
desc = f">> Validation ({epc})"
total_iteration: int = sum(len(loader) for loader in loaders) # U
total_images: int = sum(len(loader.dataset) for loader in loaders) # D
n_loss: int = max(map(len, list_loss_fns))
all_dices: Tensor = torch.zeros((total_images, K),
dtype=torch.float32,
device=device)
loss_log: Tensor = torch.zeros((total_iteration, n_loss),
dtype=torch.float32,
device=device)
three_d_dices: Optional[Tensor]
if compute_3d_dice:
three_d_dices = torch.zeros((total_iteration, K),
dtype=torch.float32,
device=device)
else:
three_d_dices = None
done_img: int = 0
done_batch: int = 0
tq_iter = tqdm_(total=total_iteration, desc=desc)
for i, (loader, loss_fns, loss_weights) in enumerate(
zip(loaders, list_loss_fns, list_loss_weights)):
for data in loader:
# t0 = time()
image: Tensor = data["images"].to(device)
target: Tensor = data["gt"].to(device)
filenames: list[str] = data["filenames"]
assert not target.requires_grad
labels: list[Tensor] = [e.to(device) for e in data["labels"]]
B, C, *_ = image.shape
# Reset gradients
if optimizer:
optimizer.zero_grad()
# Forward
pred_logits: Tensor = net(image)
pred_probs: Tensor = F.softmax(temperature * pred_logits, dim=1)
predicted_mask: Tensor = probs2one_hot(
pred_probs.detach()) # Used only for dice computation
assert not predicted_mask.requires_grad
assert len(loss_fns) == len(loss_weights) == len(labels)
ziped = zip(loss_fns, labels, loss_weights)
losses = [
w * loss_fn(pred_probs, label) for loss_fn, label, w in ziped
]
loss = reduce(add, losses)
assert loss.shape == (), loss.shape
# Backward
if optimizer:
loss.backward()
optimizer.step()
# Compute and log metrics
for j in range(len(loss_fns)):
loss_log[done_batch, j] = losses[j].detach()
sm_slice = slice(done_img,
done_img + B) # Values only for current batch
dices: Tensor = dice_coef(predicted_mask, target)
assert dices.shape == (B, K), (dices.shape, B, K)
all_dices[sm_slice, ...] = dices
if compute_3d_dice:
three_d_DSC: Tensor = dice_batch(predicted_mask, target)
assert three_d_DSC.shape == (K, )
three_d_dices[done_batch] = three_d_DSC # type: ignore
# Save images
if savedir:
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=UserWarning)
predicted_class: Tensor = probs2class(pred_probs)
save_images(predicted_class, filenames, savedir, mode, epc)
# Logging
big_slice = slice(0, done_img +
B) # Value for current and previous batches
dsc_dict: dict = {f"DSC{n}": all_dices[big_slice, n].mean() for n in metric_axis} | \
({f"3d_DSC{n}": three_d_dices[:done_batch, n].mean() for n in metric_axis}
if three_d_dices is not None else {})
loss_dict = {
f"loss_{i}": loss_log[:done_batch].mean(dim=0)[i]
for i in range(n_loss)
}
stat_dict = dsc_dict | loss_dict
nice_dict = {k: f"{v:.3f}" for (k, v) in stat_dict.items()}
done_img += B
done_batch += 1
tq_iter.set_postfix({**nice_dict, "loader": str(i)})
tq_iter.update(1)
tq_iter.close()
print(f"{desc} " + ', '.join(f"{k}={v}" for (k, v) in nice_dict.items()))
return (loss_log.detach().cpu(), all_dices.detach().cpu(),
three_d_dices.detach().cpu()
if three_d_dices is not None else None)
def do_epoch(mode: str, args, net, device, use_cuda, loader, optimizer,
num_classes, epoch):
totalImages = len(loader)
if mode == "train":
net.train()
desc = f">> Training ({epoch})"
elif mode == "val":
net.eval()
desc = f">> Validation ({epoch})"
total_iteration, total_images = len(loader), len(loader.dataset)
all_dices: Tensor = torch.zeros((total_images, num_classes),
dtype=eval(args.dtype),
device=device)
batch_dices: Tensor = torch.zeros((total_iteration, num_classes),
dtype=eval(args.dtype),
device=device)
loss_log: Tensor = torch.zeros((total_images),
dtype=eval(args.dtype),
device=device)
entropy_log: Tensor = torch.zeros((total_images),
dtype=eval(args.dtype),
device=device)
KL_log: Tensor = torch.zeros((total_images),
dtype=eval(args.dtype),
device=device)
tq_iter = tqdm_(enumerate(loader), total=total_iteration, desc=desc)
done: int = 0
for j, data in tq_iter:
image_f, image_i, image_d, image_o, image_w, image_c, labels, img_names = data
#image_f=image_f.type(torch.FloatTensor)/65535.
#image_f = image_f.type(torch.FloatTensor)/65535.
#image_i = image_i.type(torch.FloatTensor)/65535.
#image_d = image_d.type(torch.FloatTensor)/65535.
#image_o = image_o.type(torch.FloatTensor)/65535.
#image_w = image_w.type(torch.FloatTensor)/65535.
#image_c = image_c.type(torch.FloatTensor)/65535.
MRI: Tensor = torch.zeros((1, 6, image_f.size()[2], image_f.size()[3]),
dtype=eval(args.dtype))
MRI = torch.cat((image_f, image_i, image_d, image_o, image_w, image_c),
dim=1)
MRI = MRI.type(
torch.FloatTensor
) / 65535.0 #.type(eval(args.dtype)) #.type(torch.FloatTensor)
targets = torch.cat((1 - labels, labels),
dim=1) #.type(torch.LongTensor)
B = len(image_f)
#print(type(labels))
#MRI = torch.cat((image_f,image_i,image_d,image_w),dim=1)
if use_cuda:
MRI, targets = MRI.to(device), targets.to(device)
# forward
outputs = net(MRI)
pred_probs = F.softmax(outputs, dim=1)
predicted_mask = probs2one_hot(pred_probs)
entropy = crossEntropy_f(pred_probs, targets)
pred_probs_aver: Tensor = torch.sum(pred_probs, dim=(2, 3))
pred_probs_aver = pred_probs_aver / torch.sum(targets).float()
target_aver: Tensor = torch.sum(targets, dim=(2, 3)).float()
target_aver = target_aver / torch.sum(targets).float()
KL_loss = args.lam * kl(target_aver, pred_probs_aver)
loss = entropy + KL_loss
if mode == "train":
# zero the parameter gradients8544
optimizer.zero_grad()
# backward + optimize
loss.backward()
optimizer.step()
# Compute and log metrics
dices: Tensor = dice_coef(predicted_mask.detach(),
targets.type(torch.cuda.IntTensor).detach())
batch_dice: Tensor = dice_batch(
predicted_mask.detach(),
targets.type(torch.cuda.IntTensor).detach())
assert batch_dice.shape == (num_classes, ) and dices.shape == (
B, num_classes), (batch_dice.shape, dices.shape, B, num_classes)
sm_slice = slice(done, done + B) # Values only for current batch
all_dices[sm_slice, ...] = dices
entropy_log[sm_slice] = entropy.detach()
loss_log[sm_slice] = loss.detach()
KL_log[sm_slice] = KL_loss.detach()
batch_dices[j] = batch_dice
# Logging
big_slice = slice(0,
done + B) # Value for current and previous batches
stat_dict = {
"dice": all_dices[big_slice, -1].mean(),
"total loss": loss_log[big_slice].mean(),
"entropy loss": entropy_log[big_slice].mean(),
"KL loss": KL_log[big_slice].mean(),
"b dice": batch_dices[:j + 1, -1].mean()
}
nice_dict = {k: f"{v:.4f}" for (k, v) in stat_dict.items()}
done += B
tq_iter.set_postfix(nice_dict)
return loss_log, entropy_log, KL_log, all_dices, batch_dices
def runInference(args: argparse.Namespace):
# print('>>> Loading the data')
# device = torch.device("cuda") if torch.cuda.is_available() and not args.cpu else torch.device("cpu")
device = torch.device("cpu")
C: int = args.num_classes
# Let's just reuse some code
png_transform = transforms.Compose([
lambda img: np.array(img)[np.newaxis, ...],
lambda nd: nd / 255, # max <= 1
lambda nd: torch.tensor(nd, dtype=torch.float32)
])
gt_transform = transforms.Compose([
lambda img: np.array(img)[np.newaxis, ...],
lambda nd: torch.tensor(nd, dtype=torch.int64),
partial(class2one_hot, C=C),
itemgetter(0)
])
bounds_gen = [(lambda *a: torch.zeros(C, 1, 2)) for _ in range(2)]
metrics = None
pred_folders = sorted(list(Path(args.pred_root).glob('iter*')))
assert len(pred_folders) == args.epochs, (len(pred_folders), args.epochs)
for epoch, pred_folder in enumerate(pred_folders):
if args.do_only and epoch not in args.do_only:
continue
# First one is dummy:
folders: List[Path] = [Path(pred_folder, 'val'), Path(pred_folder, 'val'), Path(args.gt_folder)]
names: List[str] = map_(lambda p: str(p.name), folders[0].glob("*.png"))
are_hots = [False, True, True]
# spacing_dict = pickle.load(open(Path(args.gt_folder, "..", "spacing.pkl"), 'rb'))
spacing_dict = None
dt_set = SliceDataset(names,
folders,
transforms=[png_transform, gt_transform, gt_transform],
debug=False,
C=C,
are_hots=are_hots,
in_memory=False,
spacing_dict=spacing_dict,
bounds_generators=bounds_gen,
quiet=True)
loader = DataLoader(dt_set,
num_workers=2)
# print('>>> Computing the metrics')
total_iteration, total_images = len(loader), len(loader.dataset)
if not metrics:
metrics = {"all_dices": torch.zeros((args.epochs, total_images, C), dtype=torch.float64, device=device),
"hausdorff": torch.zeros((args.epochs, total_images, C), dtype=torch.float64, device=device)}
desc = f">> Computing"
tq_iter = tqdm_(enumerate(loader), total=total_iteration, desc=desc)
done: int = 0
for j, (filenames, _, pred, gt, _) in tq_iter:
B = len(pred)
pred = pred.to(device)
gt = gt.to(device)
assert simplex(pred) and sset(pred, [0, 1])
assert simplex(gt) and sset(gt, [0, 1])
dices: Tensor = dice_coef(pred, gt)
assert dices.shape == (B, C)
haussdorf_res: Tensor = haussdorf(pred, gt)
assert haussdorf_res.shape == (B, C)
sm_slice = slice(done, done + B) # Values only for current batch
metrics["all_dices"][epoch, sm_slice, ...] = dices
metrics["hausdorff"][epoch, sm_slice, ...] = haussdorf_res
done += B
for key, v in metrics.items():
print(epoch, key, map_("{:.4f}".format, v[epoch].mean(dim=0)))
if metrics:
savedir: Path = Path(args.save_folder)
for k, e in metrics.items():
np.save(Path(savedir, f"{k}.npy"), e.cpu().numpy())
def runInference(args: argparse.Namespace):
print('>>> Loading model')
net = torch.load(args.model_weights)
device = torch.device("cuda")
net.to(device)
print('>>> Loading the data')
batch_size: int = args.batch_size
num_classes: int = args.num_classes
png_transform = transforms.Compose([
lambda img: img.convert('L'),
lambda img: np.array(img)[np.newaxis, ...],
lambda nd: nd / 255, # max <= 1
lambda nd: torch.tensor(nd, dtype=torch.float32)
])
dummy_gt = transforms.Compose([
lambda img: np.array(img), lambda nd: torch.zeros(
(num_classes, *(nd.shape)), dtype=torch.int64)
])
folders: List[Path] = [Path(args.data_folder)]
names: List[str] = map_(lambda p: str(p.name), folders[0].glob("*.png"))
dt_set = SliceDataset(
names,
folders * 2, # Duplicate for compatibility reasons
are_hots=[False, False],
transforms=[png_transform,
dummy_gt], # So it is happy about the target size
bounds_generators=[],
debug=False,
C=num_classes)
loader = DataLoader(dt_set,
batch_size=batch_size,
num_workers=batch_size + 2,
shuffle=False,
drop_last=False)
print('>>> Starting the inference')
savedir: str = args.save_folder
total_iteration = len(loader)
desc = f">> Inference"
tq_iter = tqdm_(enumerate(loader), total=total_iteration, desc=desc)
with torch.no_grad():
for j, (filenames, image, _) in tq_iter:
image = image.to(device)
pred_logits: Tensor = net(image)
pred_probs: Tensor = F.softmax(pred_logits, dim=1)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
predicted_class: Tensor
if args.mode == "argmax":
predicted_class = probs2class(pred_probs)
elif args.mode == 'probs':
predicted_class = (pred_probs[:, args.probs_class, ...] *
255).type(torch.uint8)
elif args.mode == "threshold":
thresholded: Tensor = pred_probs[:, ...] > args.threshold
predicted_class = thresholded.argmax(dim=1)
save_images(predicted_class, filenames, savedir, "", 0)
def do_epoch(args, mode: str, net: Any, device: Any, epc: int,
loss_fns: List[Callable], loss_weights: List[float],
new_w:int, C: int, metric_axis:List[int], savedir: str = "",
optimizer: Any = None, target_loader: Any = None, best_dice3d_val:Any=None):
assert mode in ["train", "val"]
L: int = len(loss_fns)
indices = torch.tensor(metric_axis,device=device)
if mode == "train":
net.train()
desc = f">> Training ({epc})"
elif mode == "val":
net.eval()
# net.train()
desc = f">> Validation ({epc})"
total_it_t, total_images_t = len(target_loader), len(target_loader.dataset)
total_iteration = total_it_t
total_images = total_images_t
if args.debug:
total_iteration = 10
pho=1
dtype = eval(args.dtype)
all_dices: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_sizes: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_sizes: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_gt_sizes: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_sizes2: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_inter_card: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_card_gt: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_card_pred: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_gt = []
all_pred = []
if args.do_hd:
all_gt: Tensor = torch.zeros((total_images, args.wh, args.wh), dtype=dtype)
all_pred: Tensor = torch.zeros((total_images, args.wh, args.wh), dtype=dtype)
loss_log: Tensor = torch.zeros((total_images), dtype=dtype, device=device)
loss_cons: Tensor = torch.zeros((total_images), dtype=dtype, device=device)
loss_se: Tensor = torch.zeros((total_images), dtype=dtype, device=device)
loss_tot: Tensor = torch.zeros((total_images), dtype=dtype, device=device)
posim_log: Tensor = torch.zeros((total_images), dtype=dtype, device=device)
haussdorf_log: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_grp: Tensor = torch.zeros((total_images, C), dtype=dtype, device=device)
all_pnames = np.zeros([total_images]).astype('U256')
#dice_3d_log: Tensor = torch.zeros((1, C), dtype=dtype, device=device)
dice_3d_log, dice_3d_sd_log = 0, 0
#dice_3d_sd_log: Tensor = torch.zeros((1, C), dtype=dtype, device=device)
hd_3d_log, asd_3d_log, hd_3d_sd_log, asd_3d_sd_log= 0, 0, 0, 0
tq_iter = tqdm_(enumerate(target_loader), total=total_iteration, desc=desc)
done: int = 0
n_warmup = args.n_warmup
mult_lw = [pho ** (epc - n_warmup + 1)] * len(loss_weights)
mult_lw[0] = 1
loss_weights = [a * b for a, b in zip(loss_weights, mult_lw)]
losses_vec, source_vec, target_vec, baseline_target_vec = [], [], [], []
pen_count = 0
with warnings.catch_warnings():
warnings.simplefilter("ignore")
count_losses = 0
for j, target_data in tq_iter:
target_data[1:] = [e.to(device) for e in target_data[1:]] # Move all tensors to device
filenames_target, target_image, target_gt = target_data[:3]
#print("target", filenames_target)
labels = target_data[3:3+L]
bounds = target_data[3+L:]
filenames_target = [f.split('.nii')[0] for f in filenames_target]
assert len(labels) == len(bounds), len(bounds)
B = len(target_image)
# Reset gradients
if optimizer:
optimizer.zero_grad()
# Forward
with torch.set_grad_enabled(mode == "train"):
pred_logits: Tensor = net(target_image)
pred_probs: Tensor = F.softmax(pred_logits, dim=1)
if new_w > 0:
pred_probs = resize(pred_probs, new_w)
labels = [resize(label, new_w) for label in labels]
target = resize(target, new_w)
predicted_mask: Tensor = probs2one_hot(pred_probs) # Used only for dice computation
assert len(bounds) == len(loss_fns) == len(loss_weights)
if epc < n_warmup:
loss_weights = [0]*len(loss_weights)
loss: Tensor = torch.zeros(1, requires_grad=True).to(device)
loss_vec = []
loss_kw = []
for loss_fn,label, w, bound in zip(loss_fns,labels, loss_weights, bounds):
if w > 0:
if eval(args.target_losses)[0][0]=="EntKLProp":
loss_1, loss_cons_prior,est_prop = loss_fn(pred_probs, label, bound)
loss = loss_1 + loss_cons_prior
else:
loss = loss_fn(pred_probs, label, bound)
loss = w*loss
loss_1 = loss
loss_kw.append(loss_1.detach())
# Backward
if optimizer:
loss.backward()
optimizer.step()
dices, inter_card, card_gt, card_pred = dice_coef(predicted_mask.detach(), target_gt.detach())
assert dices.shape == (B, C), (dices.shape, B, C)
sm_slice = slice(done, done + B) # Values only for current batch
all_dices[sm_slice, ...] = dices
if eval(args.target_losses)[0][0] in ["EntKLProp"]:
all_sizes[sm_slice, ...] = torch.round(est_prop.detach()*target_image.shape[2]*target_image.shape[3])
all_sizes2[sm_slice, ...] = torch.sum(predicted_mask,dim=(2,3))
all_gt_sizes[sm_slice, ...] = torch.sum(target_gt,dim=(2,3))
all_grp[sm_slice, ...] = torch.FloatTensor(get_subj_nb(filenames_target)).unsqueeze(1).repeat(1,C)
all_pnames[sm_slice] = filenames_target
all_inter_card[sm_slice, ...] = inter_card
all_card_gt[sm_slice, ...] = card_gt
all_card_pred[sm_slice, ...] = card_pred
if args.do_hd:
all_pred[sm_slice, ...] = probs2class(predicted_mask[:,:,:,:]).cpu().detach()
all_gt[sm_slice, ...] = probs2class(target_gt).detach()
loss_se[sm_slice] = loss_kw[0]
if len(loss_kw)>1:
loss_cons[sm_slice] = loss_kw[1]
loss_tot[sm_slice] = loss_kw[1]+loss_kw[0]
else:
loss_cons[sm_slice] = 0
loss_tot[sm_slice] = loss_kw[0]
# # Save images
if savedir and args.saveim and mode =="val":
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=UserWarning)
warnings.simplefilter("ignore")
predicted_class: Tensor = probs2class(pred_probs)
save_images(predicted_class, filenames_target, savedir, mode, epc, False)
if args.entmap:
ent_map = torch.einsum("bcwh,bcwh->bwh", [-pred_probs, (pred_probs+1e-10).log()])
save_images_ent(ent_map, filenames_target, savedir,'ent_map', epc)
# Logging
big_slice = slice(0, done + B) # Value for current and previous batches
stat_dict = {**{f"DSC{n}": all_dices[big_slice, n].mean() for n in metric_axis},
**{f"SZ{n}": all_sizes[big_slice, n].mean() for n in metric_axis},
**({f"DSC_source{n}": all_dices_s[big_slice, n].mean() for n in metric_axis}
if args.source_metrics else {})}
size_dict = {**{f"SZ{n}": all_sizes[big_slice, n].mean() for n in metric_axis}}
nice_dict = {k: f"{v:.4f}" for (k, v) in stat_dict.items()}
done += B
tq_iter.set_postfix(nice_dict)
if args.dice_3d and (mode == 'val'):
dice_3d_log, dice_3d_sd_log,asd_3d_log, asd_3d_sd_log,hd_3d_log, hd_3d_sd_log = dice3d(all_grp, all_inter_card, all_card_gt, all_card_pred,all_pred,all_gt,all_pnames,metric_axis,args.pprint,args.do_hd,args.do_asd,best_dice3d_val)
dice_2d = torch.index_select(all_dices, 1, indices).mean().cpu().numpy().item()
target_vec = [dice_3d_log, dice_3d_sd_log,asd_3d_log, asd_3d_sd_log,hd_3d_log,hd_3d_sd_log,dice_2d]
size_mean = torch.index_select(all_sizes2, 1, indices).mean(dim=0).cpu().numpy()
size_gt_mean = torch.index_select(all_gt_sizes, 1, indices).mean(dim=0).cpu().numpy()
mask_pos = torch.index_select(all_sizes2, 1, indices)!=0
gt_pos = torch.index_select(all_gt_sizes, 1, indices)!=0
size_mean_pos = torch.index_select(all_sizes2, 1, indices).sum(dim=0).cpu().numpy()/mask_pos.sum(dim=0).cpu().numpy()
gt_size_mean_pos = torch.index_select(all_gt_sizes, 1, indices).sum(dim=0).cpu().numpy()/gt_pos.sum(dim=0).cpu().numpy()
size_mean2 = torch.index_select(all_sizes2, 1, indices).mean(dim=0).cpu().numpy()
losses_vec = [loss_se.mean().item(),loss_cons.mean().item(),loss_tot.mean().item(),size_mean.mean(),size_mean_pos.mean(),size_gt_mean.mean(),gt_size_mean_pos.mean()]
if not epc%50:
df_t = pd.DataFrame({
"val_ids":all_pnames,
"proposal_size":all_sizes2.cpu()})
df_t.to_csv(Path(savedir,mode+str(epc)+"sizes.csv"), float_format="%.4f", index_label="epoch")
return losses_vec, target_vec,source_vec
def do_epoch(args, mode: str, net: Any, device: Any, loader: DataLoader, epc: int,
loss_fns: List[Callable], loss_weights: List[float],loss_fns_source: List[Callable],
loss_weights_source: List[float], new_w:int, num_steps:int, C: int, savedir: str = "",
optimizer: Any = None, target_loader: Any = None, lambda_adv_target:float =0.001) -> Tuple[List,List,List,List]:
assert mode in ["train", "val"]
L: int = len(loss_fns)
if mode == "train":
net.train()
desc = f">> Training ({epc})"
elif mode == "val":
net.eval()
desc = f">> Validation ({epc})"
total_iteration, total_images = len(loader), len(loader.dataset)
# losses metrics
loss_seg_log = np.zeros(total_images)
loss_cons_log = np.zeros(total_images)
loss_inf_log = np.zeros(total_images)
loss_adv_log = np.zeros(total_images)
loss_D_log = np.zeros(total_images)
# source metrics
dices_log_s = np.zeros((total_images, C))
posim_log_s = np.zeros(total_images)
haussdorf_log_s = np.zeros((total_images, C))
# target metrics
dices_log_t = np.zeros((total_images, C))
dices_baseline_log_t = np.zeros((total_images, C))
posim_log_t = np.zeros(total_images)
haussdorf_log_t = np.zeros((total_images, C))
cudnn.benchmark = True
model_D = FCDiscriminator(num_classes=C)
model_D.train()
model_D.to(device)
optimizer_D = torch.optim.Adam(model_D.parameters(), lr=args.l_rate_D, betas=(0.9, 0.99))
tq_iter = tqdm_(enumerate(zip(loader, target_loader)), total=total_iteration, desc=desc)
done: int = 0
dice_3d_s = 0
dice_3d_sd_s = 0
dice_3d_t = 0
dice_3d_sd_t = 0
baseline_target_vec = [0,0]
with warnings.catch_warnings():
warnings.simplefilter("ignore")
for j, (source_data, target_data) in tq_iter:
source_data[1:] = [e.to(device) for e in source_data[1:]] # Move all tensors to device
filenames_source, source_image, source_gt = source_data
target_data[1:] = [e.to(device) for e in target_data[1:]] # Move all tensors to device
filenames_target, target_image, target_gt = target_data[:3]
labels = target_data[3:3+L]
bounds = target_data[3+L:]
assert len(labels) == len(bounds)
B = len(target_image)
#print("source: %s , target: %s" % (filenames_source, filenames_target))
source_probs, target_probs, loss, loss_seg, loss_adv, loss_inf, loss_cons, loss_D = for_back_step_comb(optimizer, mode, source_image, target_image, source_gt, labels,
net, loss_fns, loss_weights,loss_fns_source, loss_weights_source, new_w, device, bounds,
model_D, optimizer_D, lambda_adv_target)
#compute metrics for current batch
if new_w > 0:
source_gt = resize(source_gt, new_w)
target_gt = resize(target_gt, new_w)
labels[0] = resize(labels[0], new_w)
dices_s, _, posim_s, haussdorf_s = compute_metrics(source_probs, source_gt, source_gt)
dices_t, dices_baseline_t, posim_t, haussdorf_t = compute_metrics(target_probs, target_gt, labels[0])
do_save_images(target_probs, savedir, filenames_target, mode, epc)
do_save_images(source_probs, savedir, filenames_source, "_".join(("source", mode)), epc)
# keep metrics in ndarrays
sm_slice = slice(done, done + B)
loss_seg_log[sm_slice] = loss_seg
loss_cons_log[sm_slice] = loss_cons
loss_adv_log[sm_slice] = loss_adv
loss_inf_log[sm_slice] = loss_inf
loss_D_log[sm_slice] = loss_D
dices_log_s[sm_slice, ...] = dices_s
haussdorf_log_s[sm_slice] = haussdorf_s
posim_log_s[sm_slice] = posim_s
dices_log_t[sm_slice, ...] = dices_t
dices_baseline_log_t[sm_slice, ...] = dices_baseline_t
haussdorf_log_t[sm_slice] = haussdorf_t
posim_log_t[sm_slice] = posim_t
done +=B
# calculate mean of metrics on all images
loss_seg_log = loss_seg_log.mean()
loss_adv_log = loss_adv_log.mean()
loss_cons_log = loss_cons_log.mean()
loss_inf_log = loss_inf_log.mean()
loss_D_log = loss_inf_log.mean()
# first select positive and negative images
dice_posim_log_s = np.compress(posim_log_s,[dices_log_s[:,1]]).mean()
dice_negim_log_s = np.compress(1-posim_log_s, [dices_log_s[:,1]]).mean()
dice_posim_log_t = np.compress(posim_log_t, [dices_log_t[:,1]]).mean()
dice_negim_log_t = np.compress(1-posim_log_t, [dices_log_t[:,1]]).mean()
# mean on the source images
dices_log_s = dices_log_s[:, -1].mean()
haussdorf_log_s = haussdorf_log_s[:, -1].mean()
# mean on the target images
dices_log_t = dices_log_t[:, -1].mean()
haussdorf_log_t = haussdorf_log_t[:, -1].mean()
# dice3D gives back the 3d dice mean on images
if not args.debug:
dice_3d_s, dice_3d_sd_s = dice3d(args.workdir, f"iter{epc:03d}", "source_"+mode, "Subj_\\d+_", args.dataset+mode+'/GT')
dice_3d_t, dice_3d_sd_t = dice3d(args.workdir, f"iter{epc:03d}", mode, "Subj_\\d+_", args.target_dataset+mode+'/GT')
if epc == 0:
dice_3d_baseline, dice_3d_sd_baseline = dice3d(args.target_dataset, mode, 'Wat_on_Inn_n', "Subj_\\d+_",args.target_dataset+mode+'/GT')
baseline_target_vec = [dice_3d_baseline, dice_3d_sd_baseline]
stat_dict = {"dice 3D source": dice_3d_s,
"dice 3D target": dice_3d_t}
nice_dict = {k: f"{v:.4f}" for (k, v) in stat_dict.items()}
print(f"{desc} " + ', '.join(f"{k}={v}" for (k, v) in nice_dict.items()))
# Keep metrics in vectors
losses_vec = [loss_seg_log, loss_adv_log,loss_inf_log, loss_cons_log , loss_D_log]
source_vec = [dices_log_s, dice_posim_log_s, dice_negim_log_s, dice_3d_s, dice_3d_sd_s, haussdorf_log_s]
target_vec = [dices_log_t, dice_posim_log_t, dice_negim_log_t, dice_3d_t, dice_3d_sd_t, haussdorf_log_t]
return losses_vec, source_vec, target_vec, baseline_target_vec
def do_epoch(
mode: str,
net: Any,
device: Any,
loaders: List[DataLoader],
epc: int,
list_loss_fns: List[List[Callable]],
list_loss_weights: List[List[float]],
K: int,
savedir: str = "",
optimizer: Any = None,
metric_axis: List[int] = [1],
compute_hausdorff: bool = False,
compute_miou: bool = False,
compute_3d_dice: bool = False,
temperature: float = 1
) -> Tuple[Tensor, Tensor, Optional[Tensor], Optional[Tensor],
Optional[Tensor]]:
assert mode in ["train", "val"]
if mode == "train":
net.train()
desc = f">> Training ({epc})"
elif mode == "val":
net.eval()
desc = f">> Validation ({epc})"
total_iteration: int = sum(len(loader) for loader in loaders) # U
total_images: int = sum(len(loader.dataset) for loader in loaders) # D
n_loss: int = max(map(len, list_loss_fns))
all_dices: Tensor = torch.zeros((total_images, K),
dtype=torch.float32,
device=device)
loss_log: Tensor = torch.zeros((total_iteration, n_loss),
dtype=torch.float32,
device=device)
iiou_log: Optional[Tensor]
intersections: Optional[Tensor]
unions: Optional[Tensor]
if compute_miou:
iiou_log = torch.zeros((total_images, K),
dtype=torch.float32,
device=device)
intersections = torch.zeros((total_images, K),
dtype=torch.float32,
device=device)
unions = torch.zeros((total_images, K),
dtype=torch.float32,
device=device)
else:
iiou_log = None
intersections = None
unions = None
three_d_dices: Optional[Tensor]
if compute_3d_dice:
three_d_dices = torch.zeros((total_iteration, K),
dtype=torch.float32,
device=device)
else:
three_d_dices = None
hausdorff_log: Optional[Tensor]
if compute_hausdorff:
hausdorff_log = torch.zeros((total_images, K),
dtype=torch.float32,
device=device)
else:
hausdorff_log = None
few_axis: bool = len(metric_axis) <= 3
done_img: int = 0
done_batch: int = 0
tq_iter = tqdm_(total=total_iteration, desc=desc)
for i, (loader, loss_fns, loss_weights) in enumerate(
zip(loaders, list_loss_fns, list_loss_weights)):
for data in loader:
image: Tensor = data["images"].to(device)
target: Tensor = data["gt"].to(device)
spacings: Tensor = data["spacings"] # Keep that one on CPU
assert not target.requires_grad
labels: List[Tensor] = [e.to(device) for e in data["labels"]]
bounds: List[Tensor] = [e.to(device) for e in data["bounds"]]
box_priors: List[List[Tuple[
Tensor, Tensor]]] # one more level for the batch
box_priors = [[(m.to(device), b.to(device)) for (m, b) in B]
for B in data["box_priors"]]
assert len(labels) == len(bounds)
B, C, *_ = image.shape
samplings: List[List[Tuple[slice]]] = data["samplings"]
assert len(samplings) == B
assert len(samplings[0][0]) == len(
image[0, 0].shape), (samplings[0][0], image[0, 0].shape)
probs_receptacle: Tensor = -torch.ones_like(
target, dtype=torch.float32) # -1 for unfilled
mask_receptacle: Tensor = -torch.ones_like(
target, dtype=torch.int32) # -1 for unfilled
# Use the sampling coordinates of the first batch item
assert not (len(samplings[0]) > 1 and
B > 1), samplings # No subsampling if batch size > 1
loss_sub_log: Tensor = torch.zeros(
(len(samplings[0]), len(loss_fns)),
dtype=torch.float32,
device=device)
for k, sampling in enumerate(samplings[0]):
img_sampling = [slice(0, B), slice(0, C)] + list(sampling)
label_sampling = [slice(0, B), slice(0, K)] + list(sampling)
assert len(img_sampling) == len(image.shape), (img_sampling,
image.shape)
sub_img = image[img_sampling]
# Reset gradients
if optimizer:
optimizer.zero_grad()
# Forward
pred_logits: Tensor = net(sub_img)
pred_probs: Tensor = F.softmax(temperature * pred_logits,
dim=1)
predicted_mask: Tensor = probs2one_hot(
pred_probs.detach()) # Used only for dice computation
assert not predicted_mask.requires_grad
probs_receptacle[label_sampling] = pred_probs[...]
mask_receptacle[label_sampling] = predicted_mask[...]
assert len(bounds) == len(loss_fns) == len(
loss_weights) == len(labels)
ziped = zip(loss_fns, labels, loss_weights, bounds)
losses = [
w * loss_fn(pred_probs, label[label_sampling], bound,
box_priors)
for loss_fn, label, w, bound in ziped
]
loss = reduce(add, losses)
assert loss.shape == (), loss.shape
# Backward
if optimizer:
loss.backward()
optimizer.step()
# Compute and log metrics
for j in range(len(loss_fns)):
loss_sub_log[k, j] = losses[j].detach()
reduced_loss_sublog: Tensor = loss_sub_log.sum(dim=0)
assert reduced_loss_sublog.shape == (len(loss_fns), ), (
reduced_loss_sublog.shape, len(loss_fns))
loss_log[done_batch, ...] = reduced_loss_sublog[...]
del loss_sub_log
sm_slice = slice(done_img,
done_img + B) # Values only for current batch
dices: Tensor = dice_coef(mask_receptacle, target)
assert dices.shape == (B, K), (dices.shape, B, K)
all_dices[sm_slice, ...] = dices
if compute_3d_dice:
three_d_DSC: Tensor = dice_batch(mask_receptacle, target)
assert three_d_DSC.shape == (K, )
three_d_dices[done_batch] = three_d_DSC # type: ignore
if compute_hausdorff:
hausdorff_res: Tensor
try:
hausdorff_res = hausdorff(mask_receptacle, target,
spacings)
except RuntimeError:
hausdorff_res = torch.zeros((B, K), device=device)
assert hausdorff_res.shape == (B, K)
hausdorff_log[sm_slice] = hausdorff_res # type: ignore
if compute_miou:
IoUs: Tensor = iIoU(mask_receptacle, target)
assert IoUs.shape == (B, K), IoUs.shape
iiou_log[sm_slice] = IoUs # type: ignore
intersections[sm_slice] = inter_sum(mask_receptacle,
target) # type: ignore
unions[sm_slice] = union_sum(mask_receptacle,
target) # type: ignore
# if False and target[0, 1].sum() > 0: # Useful template for quick and dirty inspection
# import matplotlib.pyplot as plt
# from pprint import pprint
# from mpl_toolkits.axes_grid1 import ImageGrid
# from utils import soft_length
# print(data["filenames"])
# pprint(data["bounds"])
# pprint(soft_length(mask_receptacle))
# fig = plt.figure()
# fig.clear()
# grid = ImageGrid(fig, 211, nrows_ncols=(1, 2))
# grid[0].imshow(data["images"][0, 0], cmap="gray")
# grid[0].contour(data["gt"][0, 1], cmap='jet', alpha=.75, linewidths=2)
# grid[1].imshow(data["images"][0, 0], cmap="gray")
# grid[1].contour(mask_receptacle[0, 1], cmap='jet', alpha=.75, linewidths=2)
# plt.show()
# Save images
if savedir:
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=UserWarning)
predicted_class: Tensor = probs2class(pred_probs)
save_images(predicted_class, data["filenames"], savedir,
mode, epc)
# Logging
big_slice = slice(0, done_img +
B) # Value for current and previous batches
dsc_dict: Dict
if few_axis:
dsc_dict = {
**{
f"DSC{n}": all_dices[big_slice, n].mean()
for n in metric_axis
},
**({
f"3d_DSC{n}": three_d_dices[:done_batch, n].mean()
for n in metric_axis
} if three_d_dices is not None else {})
}
else:
dsc_dict = {}
# dsc_dict = {f"DSC{n}": all_dices[big_slice, n].mean() for n in metric_axis} if few_axis else {}
hauss_dict = {f"HD{n}": hausdorff_log[big_slice, n].mean() for n in metric_axis} \
if hausdorff_log is not None and few_axis else {}
miou_dict = {f"iIoU": iiou_log[big_slice, metric_axis].mean(),
f"mIoU": (intersections.sum(dim=0) / (unions.sum(dim=0) + 1e-10)).mean()} \
if iiou_log is not None and intersections is not None and unions is not None else {}
if len(metric_axis) > 1:
mean_dict = {"DSC": all_dices[big_slice, metric_axis].mean()}
if hausdorff_log:
mean_dict["HD"] = hausdorff_log[big_slice,
metric_axis].mean()
else:
mean_dict = {}
stat_dict = {
**miou_dict,
**dsc_dict,
**hauss_dict,
**mean_dict, "loss": loss_log[:done_batch].mean()
}
nice_dict = {k: f"{v:.3f}" for (k, v) in stat_dict.items()}
done_img += B
done_batch += 1
tq_iter.set_postfix({**nice_dict, "loader": str(i)})
tq_iter.update(1)
tq_iter.close()
print(f"{desc} " + ', '.join(f"{k}={v}" for (k, v) in nice_dict.items()))
mIoUs: Optional[Tensor]
if intersections and unions:
mIoUs = (intersections.sum(dim=0) / (unions.sum(dim=0) + 1e-10))
assert mIoUs.shape == (K, ), mIoUs.shape
else:
mIoUs = None
if not few_axis and False:
print(f"DSC: {[f'{all_dices[:, n].mean():.3f}' for n in metric_axis]}")
print(f"iIoU: {[f'{iiou_log[:, n].mean():.3f}' for n in metric_axis]}")
if mIoUs:
print(f"mIoU: {[f'{mIoUs[n]:.3f}' for n in metric_axis]}")
return (
loss_log.detach().cpu(), all_dices.detach().cpu(),
hausdorff_log.detach().cpu() if hausdorff_log is not None else None,
mIoUs.detach().cpu() if mIoUs is not None else None,
three_d_dices.detach().cpu() if three_d_dices is not None else None)
def do_epoch(mode: str, net: Any, device: Any, loaders: list[DataLoader], epc: int,
list_loss_fns: list[list[Callable]], list_loss_weights: list[list[float]], K: int,
savedir: Path = None, optimizer: Any = None,
metric_axis: list[int] = [1], requested_metrics: list[str] = None,
temperature: float = 1) -> dict[str, Tensor]:
assert mode in ["train", "val", "dual"]
if requested_metrics is None:
requested_metrics = []
if mode == "train":
net.train()
desc = f">> Training ({epc})"
elif mode == "val":
net.eval()
desc = f">> Validation ({epc})"
elif mode == "dual":
net.eval()
desc = f">> Dual ({epc})"
total_iteration: int = sum(len(loader) for loader in loaders) # U
total_images: int = sum(len(loader.dataset) for loader in loaders) # D
n_loss: int = max(map(len, list_loss_fns))
epoch_metrics: dict[str, Tensor]
epoch_metrics = {"dice": torch.zeros((total_images, K), dtype=torch.float32, device=device),
"loss": torch.zeros((total_iteration, n_loss), dtype=torch.float32, device=device)}
if "3d_dsc" in requested_metrics:
epoch_metrics["3d_dsc"] = torch.zeros((total_iteration, K), dtype=torch.float32, device=device)
few_axis: bool = len(metric_axis) <= 4
# time_log: np.ndarray = np.ndarray(total_iteration, dtype=np.float32)
done_img: int = 0
done_batch: int = 0
tq_iter = tqdm_(total=total_iteration, desc=desc)
for i, (loader, loss_fns, loss_weights) in enumerate(zip(loaders, list_loss_fns, list_loss_weights)):
for data in loader:
# t0 = time()
image: Tensor = data["images"].to(device)
target: Tensor = data["gt"].to(device)
filenames: list[str] = data["filenames"]
assert not target.requires_grad
labels: list[Tensor] = [e.to(device) for e in data["labels"]]
bounds: list[Tensor] = [e.to(device) for e in data["bounds"]]
assert len(labels) == len(bounds)
B, C, *_ = image.shape
samplings: list[list[Tuple[slice]]] = data["samplings"]
assert len(samplings) == B
assert len(samplings[0][0]) == len(image[0, 0].shape), (samplings[0][0], image[0, 0].shape)
probs_receptacle: Tensor = - torch.ones_like(target, dtype=torch.float32) # -1 for unfilled
mask_receptacle: Tensor = - torch.ones_like(target, dtype=torch.int32) # -1 for unfilled
# Use the sampling coordinates of the first batch item
assert not (len(samplings[0]) > 1 and B > 1), samplings # No subsampling if batch size > 1
loss_sub_log: Tensor = torch.zeros((len(samplings[0]), len(loss_fns)),
dtype=torch.float32, device=device)
for k, sampling in enumerate(samplings[0]):
img_sampling = [slice(0, B), slice(0, C)] + list(sampling)
label_sampling = [slice(0, B), slice(0, K)] + list(sampling)
assert len(img_sampling) == len(image.shape), (img_sampling, image.shape)
sub_img = image[img_sampling]
# Reset gradients
if optimizer:
optimizer.zero_grad()
# Forward
pred_logits: Tensor = net(sub_img)
pred_probs: Tensor = F.softmax(temperature * pred_logits, dim=1)
# Used only for dice computation:
predicted_mask: Tensor = probs2one_hot(pred_probs.detach())
assert not predicted_mask.requires_grad
probs_receptacle[label_sampling] = pred_probs[...]
mask_receptacle[label_sampling] = predicted_mask[...]
assert len(bounds) == len(loss_fns) == len(loss_weights) == len(labels)
ziped = zip(loss_fns, labels, loss_weights, bounds)
losses = [w * loss_fn(pred_probs, label[label_sampling], bound, filenames)
for loss_fn, label, w, bound in ziped]
loss = reduce(add, losses)
assert loss.shape == (), loss.shape
# Backward
if optimizer:
loss.backward()
optimizer.step()
# Compute and log metrics
for j in range(len(loss_fns)):
loss_sub_log[k, j] = losses[j].detach()
reduced_loss_sublog: Tensor = loss_sub_log.sum(dim=0)
assert reduced_loss_sublog.shape == (len(loss_fns),), (reduced_loss_sublog.shape, len(loss_fns))
epoch_metrics["loss"][done_batch, ...] = reduced_loss_sublog[...]
del loss_sub_log
sm_slice = slice(done_img, done_img + B) # Values only for current batch
dices: Tensor = dice_coef(mask_receptacle, target)
assert dices.shape == (B, K), (dices.shape, B, K)
epoch_metrics["dice"][sm_slice, ...] = dices
if "3d_dsc" in requested_metrics:
three_d_DSC: Tensor = dice_batch(mask_receptacle, target)
assert three_d_DSC.shape == (K,)
epoch_metrics["3d_dsc"][done_batch] = three_d_DSC # type: ignore
# Save images
if savedir:
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=UserWarning)
predicted_class: Tensor = probs2class(pred_probs)
save_images(predicted_class,
data["filenames"],
savedir / f"iter{epc:03d}" / mode)
# Logging
big_slice = slice(0, done_img + B) # Value for current and previous batches
stat_dict: dict[str, Any] = {}
# The order matters for the final display -- it is easy to change
if few_axis:
stat_dict |= {f"DSC{n}": epoch_metrics["dice"][big_slice, n].mean()
for n in metric_axis}
if "3d_dsc" in requested_metrics:
stat_dict |= {f"3d_DSC{n}": epoch_metrics["3d_dsc"][:done_batch, n].mean()
for n in metric_axis}
if len(metric_axis) > 1:
stat_dict |= {"DSC": epoch_metrics["dice"][big_slice, metric_axis].mean()}
stat_dict |= {f"loss_{i}": epoch_metrics["loss"][:done_batch].mean(dim=0)[i]
for i in range(n_loss)}
nice_dict = {k: f"{v:.3f}" for (k, v) in stat_dict.items()}
# t1 = time()
# time_log[done_batch] = (t1 - t0)
done_img += B
done_batch += 1
tq_iter.set_postfix({**nice_dict, "loader": str(i)})
tq_iter.update(1)
tq_iter.close()
print(f"{desc} " + ', '.join(f"{k}={v}" for (k, v) in nice_dict.items()))
return {k: v.detach().cpu() for (k, v) in epoch_metrics.items()}
