pre.append(int(argmax(pred, 0)))
else:
pre += argmax(pred, 1).tolist()
pbar.update(img.shape[0])
return np.array(pre)
df = pd.DataFrame(columns=[
'filename', 'samples', 'class0', 'class1', 'class2', 'class3', 'gt'
])
tifs = os.listdir('/mnt/hd1/antoine/TissueNet/data/Train_Metadata/tif')
with tqdm(total=len(tifs), desc=f'evaluating', unit='img') as pbar2:
for tif_name in tifs:
assert (tif_name.endswith('.tif'))
tif_dir = '/mnt/hd1/antoine/TissueNet/data/Train_Metadata/tif/' + tif_name
img = np_from_tif(tif_dir, 6)
#print(img.shape)
gt = []
for i in range(4):
gt.append(int(csv[csv['filename'] == tif_name][str(i)]))
mask = seg_hsv_threshold(img,
disp=False,
denoise_f=(denoise_erode, denoise_bilatera))
#print(mask.shape)
boxes, contours, _ = get_boxes_contours(img, mask, disp=False)
R = ROI()
R.set_boxes(name=tif_name,
H=img.shape[0],
W=img.shape[1],
boxes=boxes,
coordinate='lu')
if len(list(pred.size())) == 1:
pre.append(int(argmax(pred, 0)))
else:
pre += argmax(pred, 1).tolist()
pbar.update(img.shape[0])
return np.array(pre)
with tqdm(total=len(tifs), desc=f'evaluating', unit='img') as pbar2:
for tif_name in tifs:
tif_dir = 'data/' + tif_name
layer = 8 # the layer for segmentation
img = None
while (img is None):
try:
img = np_from_tif(tif_dir, layer)
except pyvips.error.Error:
layer -= 1
else:
if img.shape[0] < 500 or img.shape[
1] < 500: # too small segmentation layer means bad resolution
img = None
layer -= 1
print('image for segmentation with size: ', img.shape)
mask = seg_hsv_threshold(img,
disp=False,
denoise_f=(denoise_erode, denoise_bilatera))
boxes, contours, _ = get_boxes_contours(img, mask, disp=False)
if len(boxes) == 0:
df = df.append([{
'filename': tif_name,
def get_segmented_patches(tif_path):
# choose propriate layer
layer = 8 # the layer for segmentation
img = None
while (img is None and layer > 0):
try:
img = np_from_tif(tif_path, layer)
except pyvips.error.Error:
layer -= 1
else:
if img.shape[0] < 500 or img.shape[
1] < 500: # too small segmentation layer means bad resolution
img = None
layer -= 1
print('image for segmentation with size: ', img.shape)
# get boxes
mask = seg_hsv_threshold(img,
disp=False,
denoise_f=(denoise_erode, denoise_bilatera))
boxes, contours, _ = get_boxes_contours(img, mask, disp=False)
if len(boxes) == 0:
return None, None, None, None
R = ROI()
R.set_boxes(name=tif_name,
H=img.shape[0],
W=img.shape[1],
boxes=boxes,
coordinate='lu')
boxes_normalized = R.get_boxes_normalized(name=tif_name, coordinate='lu')
p = 0
ind = 0
num_patches = 0
# if there are too many patches (here more than 2000), downsample the tif image
while ind == 0:
# build dataset from tif and detected ROI
if p == 0:
patches = TissueTestAnno(tif_path,
rois=boxes_normalized,
mask=mask,
stride=2048,
grid_size=2048,
page=p,
mask_page=layer,
transform=transform)
elif p == 1:
patches = TissueTestAnno(tif_path,
rois=boxes_normalized,
mask=mask,
stride=1536,
grid_size=1536,
page=p,
mask_page=layer,
transform=transform)
elif p == 2:
patches = TissueTestAnno(tif_path,
rois=boxes_normalized,
mask=mask,
stride=1024,
grid_size=1024,
page=p,
mask_page=layer,
transform=transform)
ind = 1
num_patches = len(patches)
if (num_patches < 500):
ind = 1
p += 1
if num_patches == 0:
return None, None, None, None
# initialize dataloader
val_loader = DataLoader(patches, batch_size=12, shuffle=False)
anchors_normalized = patches.get_anchors_normalized()
anchor_group_info = patches.get_anchor_group()
return val_loader, boxes_normalized, anchors_normalized, anchor_group_info