def normalize(v):
v = list2nparray(v)
_min, _max = min(v), max(v)
v -= _min
if sum(v) == 0:
return np.ones(v.shape, 'float32') / np.prod(v.shape)
return v / sum(v)
def ic01_to_i01c(arr, allow_2d=False):
"""change array of axis ('c', 0, 1) to (0, 1, 'c')"""
arr = list2nparray(arr)
if allow_2d and arr.ndim == 2:
arr = arr.reshape((1, arr.shape[0], arr.shape[1]))
assert arr.ndim == 3 and arr.shape[0] in (1, 3), arr.shape
return np.swapaxes(np.rollaxis(arr, 1), 1, 2)
def i01c_to_ic01(arr, allow_2d=False, out_4d=False):
"""change array of axis (0, 1, 'c') to ('c', 0, 1)"""
arr = list2nparray(arr)
if allow_2d and arr.ndim == 2:
arr = arr.reshape((arr.shape[0], arr.shape[1], 1))
assert arr.ndim == 3 and arr.shape[2] in (1, 3), arr.shape
arr = np.rollaxis(arr, 2)
if out_4d:
arr = arr.reshape((1, arr.shape[0], arr.shape[1], arr.shape[2]))
return arr
def histogram(img, normalize=True):
if img.ndim == 2: # grayscale
hist = cv2.calcHist([img], [0], None, [256], [0., 255.0])
if normalize:
hist = cv2.normalize(hist, hist, 0, 1, cv2.NORM_MINMAX)
return hist
else:
hists = []
for channel in xrange(img.shape[2]):
hist = cv2.calcHist([img], [channel], None, [256], [0., 255.0])
if normalize:
hist = cv2.normalize(hist, hist, 0, 1, cv2.NORM_MINMAX)
hists.append(hist)
return list2nparray(hists)
def list2nparray(lst, dtype=None):
"""fast conversion from nested list to ndarray by pre-allocating space"""
if isinstance(lst, np.ndarray):
return lst
assert isinstance(lst, (list, tuple)), 'bad type: {}'.format(type(lst))
assert lst, 'attempt to convert empty list to np array'
if isinstance(lst[0], np.ndarray):
dim1 = lst[0].shape
assert all(i.shape == dim1 for i in lst), [i.shape for i in lst]
if dtype is None:
dtype = lst[0].dtype
assert all(i.dtype == dtype for i in lst), (dtype, [i.dtype for i in lst])
elif isinstance(lst[0], (int, float, long, complex)):
return list2nparray(lst, dtype=dtype)
else:
dim1 = list2nparray(lst[0])
if dtype is None:
dtype = dim1.dtype
dim1 = dim1.shape
shape = [len(lst)] + list(dim1)
rst = np.zeros(shape, dtype=dtype)
for idx, i in enumerate(lst):
rst[idx] = i
return rst
def cosine_similarity(v0, v1):
a, b = list2nparray(v0, "float32"), list2nparray(v1, "float32")
return np.dot(a, b) / math.sqrt(sum(a ** 2)) / math.sqrt(sum(b ** 2))