def merge_masks_diff_inputs(groupkeys, batchid=''):
os.makedirs('/tmp/results/{}/inputmerged/'.format(batchid),
exist_ok=True)
masks = []
for file in groupkeys:
binarymask = cv2.resize(pngread(file), (1024, 1024),
interpolation=cv2.INTER_AREA)
distance = ndi.distance_transform_edt(binarymask)
local_maxi = peak_local_max(distance,
labels=binarymask,
footprint=np.ones((3, 3)),
indices=False)
markers = ndi.label(local_maxi)[0]
masks.append(watershed(-distance, markers, mask=binarymask))
try:
binarymask = merge_two_masks(outpaths)
distance = ndi.distance_transform_edt(binarymask)
local_maxi = peak_local_max(distance,
labels=binarymask,
footprint=np.ones((3, 3)),
indices=False)
markers = ndi.label(local_maxi)[0]
except:
mask = watershed(-distance, markers, mask=binarymask)
pass
savepath = os.path.join(
'/tmp/results/' + batchid + '/inputmerged/',
file.split('/')[-1].split('__')[0].replace(
'merged_', 'inputmerged_') + '.jpg')
pngsave(savepath, np.uint8(mask > 0))
def merge_two_masks(maskpaths):
masks = []
for mpath in maskpaths:
binarymask = pngread(mpath)
distance = ndi.distance_transform_edt(binarymask)
local_maxi = peak_local_max(distance,
labels=binarymask,
footprint=np.ones((3, 3)),
indices=False)
markers = ndi.label(local_maxi)[0]
masks.append(watershed(-distance, markers, mask=binarymask))
mask = merge_multiple_detections(masks)
return (mask)
def createmultipleinputs(inputpath):
# pad to square
im = pngread(inputpath)
if len(im.shape) == 3:
print(
'Images should be grayscale but had dimensions {} - automatically converted'
.format(im.shape))
im = np.sum(im, 2)
im = np.uint16(
img_as_int(exposure.rescale_intensity(im,
out_range=(0, 2**15 - 1))))
imshape = im.shape
edgediff = np.max(imshape) - np.min(imshape)
orig = im
if imshape[1] > imshape[0]:
orig = cv2.copyMakeBorder(im,
math.ceil(edgediff / 2),
math.ceil(edgediff / 2),
0,
0,
cv2.BORDER_CONSTANT,
value=[0, 0, 0])
if imshape[0] > imshape[1]:
orig = cv2.copyMakeBorder(im,
0,
0,
math.ceil(edgediff / 2),
math.ceil(edgediff / 2),
cv2.BORDER_CONSTANT,
value=[0, 0, 0])
# ==>resize to 1024
im1024 = cv2.resize(orig, (1024, 1024), interpolation=cv2.INTER_AREA)
# ==>resize to 720
im720 = cv2.resize(orig, (720, 720), interpolation=cv2.INTER_AREA)
# preprocess both
im1024preproc = preproc(im1024)
im720preproc = preproc(im720)
return ([orig, im1024preproc, im720preproc, im1024, im720])
def proccesshelafiles(f, s3=s3, sess=sess, bucket=bucket, prefix=prefix):
train_channel = prefix + '/train'
validation_channel = prefix + '/validation'
train_annotation_channel = prefix + '/train_annotation'
validation_annotation_channel = prefix + '/validation_annotation'
if 'jpg' in f:
file = f
if 'segproj/hela_dataset_training_data/train/' in f:
jpgpath = '/tmp/' + 'hela_' + file.split('/')[-1].split(
'.')[0] + '_' + 'raw.jpg'
s3.meta.client.download_file('meadata', file, jpgpath)
inverted_img = pngread(jpgpath)
num = int(''.join(filter(str.isdigit, str(
inverted_img.dtype)))) - 1
image = exposure.rescale_intensity(inverted_img,
out_range=(0, 2**num - 1))
image = cv2.resize(image, (1024, 1024),
interpolation=cv2.INTER_AREA)
pngsave(jpgpath, img_as_ubyte(image))
sess.upload_data(path=jpgpath,
bucket=bucket,
key_prefix=train_channel)
elif 'segproj/hela_dataset_training_data/val/' in f:
jpgpath = '/tmp/' + 'hela_' + file.split('/')[-1].split(
'.')[0] + '_' + 'raw.jpg'
s3.meta.client.download_file('meadata', file, jpgpath)
inverted_img = pngread(jpgpath)
num = int(''.join(filter(str.isdigit, str(
inverted_img.dtype)))) - 1
image = exposure.rescale_intensity(inverted_img,
out_range=(0, 2**num - 1))
image = cv2.resize(image, (1024, 1024),
interpolation=cv2.INTER_AREA)
pngsave(jpgpath, img_as_ubyte(image))
sess.upload_data(path=jpgpath,
bucket=bucket,
key_prefix=validation_channel)
elif 'png' in f:
file = f
if 'segproj/hela_dataset_training_data/train_annotation/' in f:
pngpath = '/tmp/' + 'hela_' + file.split('/')[-1].split(
'.')[0] + '_' + 'raw.png'
s3.meta.client.download_file('meadata', file, pngpath)
im1 = pngread(pngpath)
num = int(''.join(filter(str.isdigit, str(im1.dtype)))) - 1
image = exposure.rescale_intensity(im1, out_range=(0, 2**num - 1))
image = img_as_ubyte(image)
im = mark_boundaries(image,
im1,
color=[0, 0, 0],
outline_color=[0, 0, 0],
mode='outer',
background_label=0)
im2 = img_as_int(im)
im3 = np.zeros([im2.shape[0], im2.shape[1]])
im3 = im2[:, :, 0] + im2[:, :, 1] + im2[:, :, 2]
im3 = np.uint8((im3 > 0))
im3 = cv2.resize(im3, (1024, 1024), interpolation=cv2.INTER_AREA)
pngsave(pngpath, im3, check_contrast=False)
sess.upload_data(path=pngpath,
bucket=bucket,
key_prefix=train_annotation_channel)
elif 'segproj/hela_dataset_training_data/val_annotation/' in f:
pngpath = '/tmp/' + 'hela_' + file.split('/')[-1].split(
'.')[0] + '_' + 'raw.png'
s3.meta.client.download_file('meadata', file, pngpath)
im1 = pngread(pngpath)
num = int(''.join(filter(str.isdigit, str(im1.dtype)))) - 1
image = exposure.rescale_intensity(im1, out_range=(0, 2**num - 1))
image = img_as_ubyte(image)
im = mark_boundaries(image,
im1,
color=[0, 0, 0],
outline_color=[0, 0, 0],
mode='outer',
background_label=0)
im2 = img_as_int(im)
im3 = np.zeros([im2.shape[0], im2.shape[1]])
im3 = im2[:, :, 0] + im2[:, :, 1] + im2[:, :, 2]
im3 = np.uint8((im3 > 0))
im3 = cv2.resize(im3, (1024, 1024), interpolation=cv2.INTER_AREA)
pngsave(pngpath, im3, check_contrast=False)
sess.upload_data(path=pngpath,
bucket=bucket,
key_prefix=validation_annotation_channel)