def segmenter_data_transform(imb, rotate=None, normalize_pctwise=False):
if isinstance(imb, tuple) and len(imb) == 2:
imgs,labels = imb
else:
imgs = imb
# rotate image if training
if rotate is not None:
for i in xrange(imgs.shape[0]):
degrees = float(np.random.randint(rotate[0], rotate[1])) if \
isinstance(rotate, tuple) else rotate
imgs[i,0] = scipy.misc.imrotate(imgs[i,0], degrees, interp='bilinear')
if isinstance(imb, tuple):
labels[i,0] = scipy.misc.imrotate(labels[i,0], degrees, interp='bilinear')
# assume they are square
sz = c.fcn_img_size
x,y = np.random.randint(0,imgs.shape[2]-sz,2) if imgs.shape[2] > sz else (0,0)
imgs = nn.utils.floatX(imgs[:,:, x:x+sz, y:y+sz])/255.
if not normalize_pctwise:
pad = imgs.shape[2] // 5
cut = imgs[:,0,pad:-pad,pad:-pad]
mu = cut.mean(axis=(1,2)).reshape(imgs.shape[0],1,1,1)
sigma = cut.std(axis=(1,2)).reshape(imgs.shape[0],1,1,1)
imgs = (imgs - mu) / sigma
imgs = np.minimum(3, np.maximum(-3, imgs))
else:
pclow, pchigh = normalize_pctwise if isinstance(normalize_pctwise, tuple) else (20,70)
for i in xrange(imgs.shape[0]):
pl,ph = np.percentile(imgs[i],(pclow, pchigh))
imgs[i] = exposure.rescale_intensity(imgs[i], in_range=(pl, ph));
imgs[i] = 2*imgs[i]/imgs[i].max() - 1.
# or other rescaling here to approximate ~ N(0,1)
if isinstance(imb, tuple):
labels = nn.utils.floatX(labels[:,:, x:x+sz, y:y+sz])
return imgs, labels
return imgs
评论列表
文章目录
