def _build(self, img, sentinel=None, coords=None, logits=None):
"""Assume that `transform_param` has the shape of (..., n_params) where n_params = n_scales + n_shifts
and:
scale = transform_params[..., :n_scales]
shift = transform_params[..., n_scales:n_scales+n_shifts]
"""
if sentinel is not None:
raise ValueError(
'Either coords or logits must be given by kwargs!')
if coords is not None and logits is not None:
raise ValueError('Please give eithe coords or logits, not both!')
if coords is None and logits is None:
raise ValueError('Please give coords or logits!')
if coords is None:
coords = self.to_coords(logits)
axis = coords.shape.ndims - 1
sx, sy, tx, ty = tf.split(coords, 4, axis=axis)
sx, sy = (clip_preserve(s, 1e-4, s) for s in (sx, sy))
transform_params = tf.concat([sx, tx, sy, ty], -1)
if len(img.get_shape()) == 3:
img = img[..., tf.newaxis]
if len(transform_params.get_shape()) == 2:
return self._sample_image(img, transform_params)
else:
transform_params = tf.unstack(transform_params, axis=1)
samples = [self._sample_image(img, tp) for tp in transform_params]
return tf.stack(samples, axis=1)
def _build(self, img, sentinel=None, coords=None, logits=None):
"""Applies the transformation.
Either coords or logits must be given, but not both. They have to be passed as kwarg. Assume that
they have the shape of (..., n_params) where n_params = n_scales + n_shifts
and:
scale = transform_params[..., :n_scales]
shift = transform_params[..., n_scales:n_scales+n_shifts]
:param img: Tensor of images of shape `[B, H, W, C]`.
:param sentinel: Unused; it is to guard `coords` or `logits` from being passed as positional arguments.
:param coords: Tensor of coordinates in Spatial Transformer space.
:param logits: Tensor of logits; they are converted to ST space.
:return: Tensor of transformed images.
"""
if sentinel is not None:
raise ValueError(
'Either coords or logits must be given by kwargs!')
if coords is not None and logits is not None:
raise ValueError('Please give eithe coords or logits, not both!')
if coords is None and logits is None:
raise ValueError('Please give coords or logits!')
if coords is None:
coords = self.to_coords(logits)
axis = coords.shape.ndims - 1
sx, sy, tx, ty = tf.split(coords, 4, axis=axis)
sx, sy = (ops.clip_preserve(s, 1e-4, s) for s in (sx, sy))
transform_params = tf.concat([sx, tx, sy, ty], -1)
if len(img.get_shape()) == 3:
img = img[..., tf.newaxis]
if len(transform_params.get_shape()) == 2:
return self._sample_image(img, transform_params)
else:
transform_params = tf.unstack(transform_params, axis=1)
samples = [self._sample_image(img, tp) for tp in transform_params]
return tf.stack(samples, axis=1)
def tabular_kl(p, q, zero_prob_value=0., logarg_clip=None):
"""Computes KL-divergence KL(p||q) for two probability mass functions (pmf) given in a tabular form.
:param p: iterable
:param q: iterable
:param zero_prob_value: float; values below this threshold are treated as zero
:param logarg_clip: float or None, clips the argument to the log to lie in [-logarg_clip, logarg_clip] if not None
:return: iterable of brodcasted shape of (p * q), per-coordinate value of KL(p||q)
"""
p, q = (tf.cast(i, tf.float64) for i in (p, q))
non_zero = tf.greater(p, zero_prob_value)
logarg = p / q
if logarg_clip is not None:
logarg = clip_preserve(logarg, 1. / logarg_clip, logarg_clip)
log = masked_apply(logarg, tf.log, non_zero)
kl = p * log
return tf.cast(kl, tf.float32)
def log_prob(self, samples):
prob = self.prob(samples)
prob = clip_preserve(prob, 1e-16, 1.)
return tf.log(prob)