def _internal_maybe_mirror_and_pred_2D(self,
x,
num_repeats,
mirror_axes,
do_mirroring=True,
mult=None):
with torch.no_grad():
x_torch = torch.from_numpy(x).float()
if self.get_device() == "cpu":
x_torch = x_torch.cpu()
else:
x_torch = x_torch.cuda(self.get_device())
result_torch = torch.zeros([1, self.num_classes] +
list(x.shape[2:])).float()
if self.get_device() == "cpu":
result_torch = result_torch.cpu()
else:
result_torch = result_torch.cuda(self.get_device())
num_results = num_repeats
if do_mirroring:
mirror_idx = 4
num_results *= 2**len(mirror_axes)
else:
mirror_idx = 1
for i in range(num_repeats):
for m in range(mirror_idx):
if m == 0:
pred = self.inference_apply_nonlin(self(x_torch))
result_torch += 1 / num_results * pred
if m == 1 and (1 in mirror_axes):
pred = self.inference_apply_nonlin(
self(flip(x_torch, 3)))
result_torch += 1 / num_results * flip(pred, 3)
if m == 2 and (0 in mirror_axes):
pred = self.inference_apply_nonlin(
self(flip(x_torch, 2)))
result_torch += 1 / num_results * flip(pred, 2)
if m == 3 and (0 in mirror_axes) and (1 in mirror_axes):
pred = self.inference_apply_nonlin(
self(flip(flip(x_torch, 3), 2)))
result_torch += 1 / num_results * flip(
flip(pred, 3), 2)
if mult is not None:
result_torch[:, :] *= mult
return result_torch.detach().cpu().numpy()
def _internal_maybe_mirror_and_pred_2D(
self,
x: Union[np.ndarray, torch.tensor],
mirror_axes: tuple,
do_mirroring: bool = True,
mult: np.ndarray or torch.tensor = None) -> torch.tensor:
# everything in here takes place on the GPU. If x and mult are not yet on GPU this will be taken care of here
# we now return a cuda tensor! Not numpy array!
assert len(x.shape) == 4, 'x must be (b, c, x, y)'
with torch.no_grad():
x = to_cuda(maybe_to_torch(x), gpu_id=self.get_device())
result_torch = torch.zeros(
[x.shape[0], self.num_classes] + list(x.shape[2:]),
dtype=torch.float).cuda(self.get_device(), non_blocking=True)
if mult is not None:
mult = to_cuda(maybe_to_torch(mult), gpu_id=self.get_device())
if do_mirroring:
mirror_idx = 4
num_results = 2**len(mirror_axes)
else:
mirror_idx = 1
num_results = 1
for m in range(mirror_idx):
if m == 0:
pred = self.inference_apply_nonlin(self(x))
result_torch += 1 / num_results * pred
if m == 1 and (1 in mirror_axes):
pred = self.inference_apply_nonlin(self(flip(x, 3)))
result_torch += 1 / num_results * flip(pred, 3)
if m == 2 and (0 in mirror_axes):
pred = self.inference_apply_nonlin(self(flip(x, 2)))
result_torch += 1 / num_results * flip(pred, 2)
if m == 3 and (0 in mirror_axes) and (1 in mirror_axes):
pred = self.inference_apply_nonlin(
self(flip(flip(x, 3), 2)))
result_torch += 1 / num_results * flip(flip(pred, 3), 2)
if mult is not None:
result_torch[:, :] *= mult
return result_torch
def _internal_maybe_mirror_and_pred_2D(self,
x,
num_repeats,
mirror_axes,
do_mirroring=True,
mult=None):
# everything in here takes place on the GPU. If x and mult are not yet on GPU this will be taken care of here
# we now return a cuda tensor! Not numpy array!
with torch.no_grad():
x = to_cuda(maybe_to_torch(x), gpu_id=self.get_device())
mult = to_cuda(maybe_to_torch(mult), gpu_id=self.get_device())
result_torch = torch.zeros(
[x.shape[0], self.num_classes] + list(x.shape[2:]),
dtype=torch.float).cuda(self.get_device(), non_blocking=True)
num_results = num_repeats
if do_mirroring:
mirror_idx = 4
num_results *= 2**len(mirror_axes)
else:
mirror_idx = 1
for i in range(num_repeats):
for m in range(mirror_idx):
if m == 0:
pred = self.inference_apply_nonlin(self(x))
result_torch += 1 / num_results * pred
if m == 1 and (1 in mirror_axes):
pred = self.inference_apply_nonlin(self(flip(x, 3)))
result_torch += 1 / num_results * flip(pred, 3)
if m == 2 and (0 in mirror_axes):
pred = self.inference_apply_nonlin(self(flip(x, 2)))
result_torch += 1 / num_results * flip(pred, 2)
if m == 3 and (0 in mirror_axes) and (1 in mirror_axes):
pred = self.inference_apply_nonlin(
self(flip(flip(x, 3), 2)))
result_torch += 1 / num_results * flip(
flip(pred, 3), 2)
if mult is not None:
result_torch[:, :] *= mult
return result_torch
