def scale_invariant_gradient( inp, deltas=[1,2,4,8,16],
weights=[1,1,1,1,1],epsilon=0.001):
"""Computes the scale invariant gradient images
inp: Tensor
deltas: list of int
The pixel delta for the difference.
This vector must be the same length as weight.
weights: list of float
The weight factor for each difference.
This vector must be the same length as delta.
epsilon: float
epsilon value for avoiding division by zero
"""
inbatch = len(inp.shape.as_list())==4
if not inbatch:
inp = tf.expand_dims(inp,0)
inp = convert_NHWC_to_NCHW(inp)
assert len(deltas)==len(weights)
sig_images = []
for delta, weight in zip(deltas,weights):
sig_images.append(sops.scale_invariant_gradient(inp, deltas=[delta], weights=[weight], epsilon=epsilon))
output = tf.concat(sig_images,axis=1)
output = convert_NCHW_to_NHWC(output)
if not inbatch:
output = tf.squeeze(output,0)
return output
def scale_invariant_gradient(inp, deltas, weights, epsilon=0.001):
"""Computes the scale invariant gradient images
inp: Tensor
deltas: list of int
The pixel delta for the difference.
This vector must be the same length as weight.
weights: list of float
The weight factor for each difference.
This vector must be the same length as delta.
epsilon: float
epsilon value for avoiding division by zero
"""
assert len(deltas) == len(weights)
sig_images = []
for delta, weight in zip(deltas, weights):
sig_images.append(
sops.scale_invariant_gradient(inp,
deltas=[delta],
weights=[weight],
epsilon=epsilon))
return tf.concat(sig_images, axis=1)
def get_multi_sig_list(inp, sig_param_list):
""" Get multi sig with the pattern in the sig_param_list, result is concat in axis=1
inp: Tensor
sig_param_list: a list of sig_param dictionary
"""
sig_list = []
for sig_param in sig_param_list:
sig_list.append(sops.scale_invariant_gradient(inp, **sig_param))
#multi_sig = tf.concat(sig_list,axis=1)
return sig_list
def test_shape(self):
with self.test_session(use_gpu=False, force_gpu=False):
input1 = np.empty((8, 40, 31))
input2 = np.empty((8, 1, 40, 31))
input3 = np.empty((2, 2, 2, 40, 31))
inputs = (input1, input2, input3)
expected_shape = [8, 2, 40, 31]
for i in inputs:
output_tensor = ops.scale_invariant_gradient(input=i)
out_shape = output_tensor.get_shape().as_list()
self.assertAllEqual(out_shape, expected_shape)
def _test_grad(self, dtype):
#A = np.random.rand(3,4).astype(dtype)
A = np.linspace(1, 2, num=16, dtype=dtype).reshape((4, 4))
shape = A.shape
data = tf.constant(A)
output = ops.scale_invariant_gradient(input=data,
deltas=[1, 2, 4],
weights=[1, 0.5, 0.25],
epsilon=0.001)
#print(A)
#print(output.eval())
err = tf.test.compute_gradient_error(data,
shape,
output,
output.get_shape().as_list(),
x_init_value=A)
print('error', err, flush=True)
self.assertLess(err, 1e-3)