def test_hwc_to_chw():
x = np.array([[[ 1, 2, 3],
[ 4, 5, 6]
],
[[ 7, 8, 9],
[10, 11, 12]
],
[[13, 14, 15],
[16, 17, 18],
],
[[19, 20, 21],
[22, 23, 24]
]
])
assert_equals(x.shape[0], 4) # h
assert_equals(x.shape[1], 2) # w
assert_equals(x.shape[2], 3) # c
y = mu.hwc_to_chw(x)
assert_equals(y.shape[0], 3)
assert_equals(y.shape[1], 4)
assert_equals(y.shape[2], 2)
assert_equals(x[3][1][2], y[2][3][1])
for i in range(4):
for j in range(2):
for k in range(3):
assert_equals(x[i][j][k], y[k][i][j])
def read_img_caf(fpath, mean=None):
'''
load image, switch to BGR, subtract mean, and make dims C x H x W for Caffe
'''
img_dat = caffe.io.load_image(fpath) # pixel value range per channel: [0, 1]
img_dat *= 255.
# RGB to BGR
img_dat = img_dat[:, :, ::-1]
# per-channel mean subtraction
if mean is not None:
img_dat -= mean
# reorder dimensions
img_dat = mu.hwc_to_chw(img_dat)
return img_dat
def read_img_PIL(fpath, mean=None):
'''
load image, switch to BGR, subtract mean, and make dims C x H x W for Caffe
'''
img = Image.open(fpath) # pixel value range per channel: [0, 255]
img_dat = np.array(img, dtype=np.float32)
# RGB to BGR
img_dat = img_dat[:, :, ::-1]
# per-channel mean subtraction
if mean is not None:
img_dat -= mean
# reorder dimensions
img_dat = mu.hwc_to_chw(img_dat)
return img_dat
def read_img_PIL(fpath, mean=None):
'''
load image, switch to BGR, subtract mean, and make dims C x H x W for Caffe
'''
img = Image.open(fpath) # pixel value range per channel: [0, 255]
img_dat = np.array(img, dtype=np.float32)
# RGB to BGR
img_dat = img_dat[:, :, ::-1]
# per-channel mean subtraction
if mean is not None:
img_dat -= mean
# reorder dimensions
img_dat = mu.hwc_to_chw(img_dat)
return img_dat
def read_img_cv2(fpath, mean=None):
'''
load image in BGR, subtract mean, and make dims C x H x W for Caffe
'''
img_dat = cv2.imread(fpath) # pixel value range per channel: [0, 255]
# channels already in BGR order
# per-channel mean subtraction
if mean is not None:
img_dat = img_dat.astype(np.float32)
img_dat -= mean
img_dat = img_dat.astype(np.float)
# reorder dimensions
img_dat = mu.hwc_to_chw(img_dat)
# casting to np.float enables plugging into protobuf
return img_dat
def read_img_caf(fpath, mean=None):
'''
load image, switch to BGR, subtract mean, and make dims C x H x W for Caffe
'''
img_dat = caffe.io.load_image(
fpath) # pixel value range per channel: [0, 1]
img_dat *= 255.
# RGB to BGR
img_dat = img_dat[:, :, ::-1]
# per-channel mean subtraction
if mean is not None:
img_dat -= mean
# reorder dimensions
img_dat = mu.hwc_to_chw(img_dat)
return img_dat
def read_img_cv2(fpath, mean=None):
'''
load image in BGR, subtract mean, and make dims C x H x W for Caffe
'''
img_dat = cv2.imread(fpath) # pixel value range per channel: [0, 255]
# channels already in BGR order
# per-channel mean subtraction
if mean is not None:
img_dat = img_dat.astype(np.float32)
img_dat -= mean
img_dat = img_dat.astype(np.float)
# reorder dimensions
img_dat = mu.hwc_to_chw(img_dat)
# casting to np.float enables plugging into protobuf
return img_dat
