def plot_img_and_seg(
img: torch.Tensor,
seg: torch.Tensor,
figsize: Optional[Tuple[float, float]] = None,
experiment: Optional[neptune.experiments.Experiment] = None,
neptune_name: Optional[str] = None):
_exp = experiment if experiment else neptune
assert len(img.shape) == len(seg.shape) == 4
n_row = img.shape[-4]
if n_row <= 1:
fig, axes = plt.subplots(ncols=2, figsize=figsize)
axes[0].imshow(img[0, 0], cmap='gray')
axes[1].imshow(seg[0, 0], cmap='seismic', vmin=-0.5, vmax=10.5)
axes[0].set_axis_off()
axes[1].set_axis_off()
else:
fig, axes = plt.subplots(ncols=2, nrows=n_row, figsize=figsize)
for n in range(n_row):
axes[n, 0].imshow(img[n, 0], cmap='gray')
axes[n, 1].imshow(seg[n, 0], cmap='seismic', vmin=-0.5, vmax=10.5)
axes[n, 0].set_axis_off()
axes[n, 1].set_axis_off()
fig.tight_layout()
if neptune_name is not None:
#log_img_and_chart(name=neptune_name, fig=fig, experiment=experiment)
log_img_only(name=neptune_name, fig=fig, experiment=_exp)
plt.close(fig)
return fig
def plot_grid(img,
figsize: Optional[Tuple[float, float]] = None,
experiment: Optional[neptune.experiments.Experiment] = None,
neptune_name: Optional[str] = None):
_exp = experiment if experiment else neptune
assert len(img.shape) == 3
n_max = img.shape[-3]
row_max = n_max // 4
if row_max <= 1:
fig, axes = plt.subplots(ncols=n_max, figsize=figsize)
for n in range(n_max):
axes[n].imshow(img[n])
else:
fig, axes = plt.subplots(ncols=4, nrows=row_max, figsize=figsize)
for n in range(4 * row_max):
row = n // 4
col = n % 4
axes[row, col].imshow(img[n])
fig.tight_layout()
if neptune_name is not None:
#log_img_and_chart(name=neptune_name, fig=fig, experiment=experiment)
log_img_only(name=neptune_name, fig=fig, experiment=_exp)
plt.close(fig)
return fig
def plot_label_contours(
label: Union[torch.Tensor, numpy.ndarray],
image: Union[torch.Tensor, numpy.ndarray],
window: Optional[tuple] = None,
contour_thickness: int = 2,
contour_color: str = 'red',
figsize: tuple = (24, 24),
experiment: Optional[neptune.experiments.Experiment] = None,
neptune_name: Optional[str] = None):
_exp = experiment if experiment else neptune
assert len(label.shape) == 2
assert len(image.shape) == 2 or len(image.shape) == 3
assert len(label.shape) == 2
if torch.is_tensor(label):
label = label.cpu().numpy()
if torch.is_tensor(image):
if len(image.shape) == 3:
image = image.permute(1, 2, 0).cpu().numpy()
else:
image = image.cpu().numpy()
if len(image.shape) == 3 and (image.shape[-1] != 3):
image = image[..., 0]
assert image.shape[:2] == label.shape[:2]
print(window)
if window is None:
window = [0, 0, label.shape[-2], label.shape[-1]]
else:
window = (max(0, window[0]), max(0, window[1]),
min(label.shape[-2],
window[2]), min(label.shape[-1], window[3]))
contours = contours_from_labels(
label[window[0]:window[2], window[1]:window[3]], contour_thickness)
img_with_contours = draw_contours(image=image[window[0]:window[2],
window[1]:window[3]],
contours=contours,
contours_color=contour_color)
fig, ax = plt.subplots(ncols=3, figsize=figsize)
ax[0].imshow(image[window[0]:window[2], window[1]:window[3]], cmap='gray')
ax[1].imshow(img_with_contours)
ax[2].imshow(
skimage.color.label2rgb(label=label[window[0]:window[2],
window[1]:window[3]],
image=image[window[0]:window[2],
window[1]:window[3]],
alpha=0.25,
bg_label=0))
fig.tight_layout()
if neptune_name is not None:
#log_img_and_chart(name=neptune_name, fig=fig, experiment=experiment)
log_img_only(name=neptune_name, fig=fig, experiment=_exp)
plt.close(fig)
return fig
def plot_tiling(tiling,
figsize: tuple = (12, 12),
window: Optional[tuple] = None,
experiment: Optional[neptune.experiments.Experiment] = None,
neptune_name: Optional[str] = None):
_exp = experiment if experiment else neptune
if window is None:
window = [
0, 0, tiling.integer_mask.shape[-2], tiling.integer_mask.shape[-1]
]
else:
window = (max(0, window[0]), max(0, window[1]),
min(tiling.integer_mask.shape[-2],
window[2]), min(tiling.integer_mask.shape[-1],
window[3]))
fig, axes = plt.subplots(ncols=2, nrows=2, figsize=figsize)
axes[0, 0].imshow(
skimage.color.label2rgb(
label=tiling.integer_mask[0, 0, window[0]:window[2],
window[1]:window[3]].cpu().numpy(),
image=numpy.zeros_like(
tiling.integer_mask[0, 0, window[0]:window[2],
window[1]:window[3]].cpu().numpy()),
alpha=1.0,
bg_label=0))
axes[0, 1].imshow(
skimage.color.label2rgb(
label=tiling.integer_mask[0, 0, window[0]:window[2],
window[1]:window[3]].cpu().numpy(),
image=tiling.raw_image[0, 0, window[0]:window[2],
window[1]:window[3]].cpu().numpy(),
alpha=0.25,
bg_label=0))
axes[1, 0].imshow(tiling.fg_prob[0, 0, window[0]:window[2],
window[1]:window[3]].cpu().numpy(),
cmap='gray')
axes[1, 1].imshow(tiling.raw_image[0, :, window[0]:window[2],
window[1]:window[3]].cpu().permute(
1, 2, 0).squeeze(-1).numpy(),
cmap='gray')
axes[0, 0].set_title("sample integer mask")
axes[0, 1].set_title("sample integer mask")
axes[1, 0].set_title("fg prob")
axes[1, 1].set_title("raw image")
fig.tight_layout()
if neptune_name is not None:
#log_img_and_chart(name=neptune_name, fig=fig, experiment=experiment)
log_img_only(name=neptune_name, fig=fig, experiment=_exp)
plt.close(fig)
return fig
def plot_concordance(
concordance,
figsize: tuple = (12, 12),
experiment: Optional[neptune.experiments.Experiment] = None,
neptune_name: Optional[str] = None):
_exp = experiment if experiment else neptune
fig, axes = plt.subplots(figsize=figsize)
axes.imshow(concordance.intersection_mask.cpu(), cmap='gray')
axes.set_title("intersection mask, iou=" + str(concordance.iou))
fig.tight_layout()
if neptune_name is not None:
log_img_only(name=neptune_name, fig=fig, experiment=_exp)
plt.close(fig)
return fig
def show_batch(images: torch.Tensor,
n_col: int = 4,
n_padding: int = 10,
title: Optional[str] = None,
pad_value: int = 1,
normalize_range: Optional[tuple] = None,
figsize: Optional[Tuple[float, float]] = None,
experiment: Optional[neptune.experiments.Experiment] = None,
neptune_name: Optional[str] = None):
_exp = experiment if experiment else neptune
"""Visualize a torch tensor of shape: (batch x ch x width x height) """
assert len(images.shape) == 4 # batch, ch, width, height
if images.device != "cpu":
images = images.cpu()
# Always normalize the image in (0,1) either using min_max of tensor or normalize_range
grid = utils.make_grid(images,
n_col,
n_padding,
normalize=True,
range=normalize_range,
scale_each=False,
pad_value=pad_value)
fig = plt.figure(figsize=figsize)
plt.imshow(grid.detach().permute(1, 2, 0).squeeze(-1).numpy())
# plt.axis("off")
if isinstance(title, str):
plt.title(title)
fig.tight_layout()
if neptune_name is not None:
#log_img_and_chart(name=neptune_name, fig=fig, experiment=experiment)
log_img_only(name=neptune_name, fig=fig, experiment=_exp)
plt.close(fig)
return fig
conditional_crop_train = ConditionalRandomCrop(desired_w=SIZE_CROPS, desired_h=SIZE_CROPS,
min_roi_fraction=0.9, n_crops_per_image=N_TRAIN)
test_data = conditional_crop_test.crop(img=img_torch,
roi_mask=roi_mask_torch)
if torch.cuda.is_available():
print("GPU GB after defining test data ->", torch.cuda.memory_allocated()/1E9)
test_loader = SpecialDataSet(img=test_data,
store_in_cuda=False,
shuffle=False,
drop_last=False,
batch_size=BATCH_SIZE)
test_batch_example_fig = test_loader.check_batch()
log_img_only(name="test_batch_example", fig=test_batch_example_fig, experiment=exp)
train_loader = SpecialDataSet(img=img_torch,
roi_mask=roi_mask_torch,
data_augmentation=conditional_crop_train,
store_in_cuda=False,
shuffle=True,
drop_last=True,
batch_size=BATCH_SIZE)
train_batch_example_fig = train_loader.check_batch()
log_img_only(name="train_batch_example", fig=train_batch_example_fig, experiment=exp)
if torch.cuda.is_available():
print("GPU GB after train_loader ->", torch.cuda.memory_allocated()/1E9)
# Make a batch of reference images by cropping the train_data at consecutive locations
reference_imgs_list = []
