def _inverse(self, y, yy, nlf0=None, nlf1=None, iso=None, cam=None):
scale = gain_model_params(iso)
# tf.summary.histogram('fitSDN_beta1', beta1)
# tf.summary.histogram('fitSDN_beta2', beta2)
x = y
if scale is not None:
x /= scale
if self._last_layer:
return tf.layers.flatten(x)
return x
def _inverse_log_det_jacobian(self,
z,
yy,
nlf0=None,
nlf1=None,
iso=None,
cam=None):
scale = gain_model_params(iso)
# tf.summary.histogram('fitSDN_beta1', beta1)
# tf.summary.histogram('fitSDN_beta2', beta2)
if scale is None:
return tf.constant(0., dtype=z.dtype, name="ildj")
return -tf.log(scale)
def _forward_log_det_jacobian(self,
x,
yy,
nlf0=None,
nlf1=None,
iso=None,
cam=None):
if self._last_layer:
x = tf.reshape(x, (-1, self.i0, self.i1, self.ic))
# yy = tf.reshape(yy, (-1, self.i0, self.i1, self.ic))
scale = gain_model_params(iso)
if scale is None:
return tf.constant(0., dtype=x.dtype, name="fldj")
return tf.log(scale)
def _forward(self, x, yy, nlf0=None, nlf1=None, iso=None, cam=None):
if self._last_layer:
x = tf.reshape(x, (-1, self.i0, self.i1, self.ic))
# yy = tf.reshape(yy, (-1, self.i0, self.i1, self.ic))
# if yy.shape[1] == 2 * x.shape[1]: # needs squeezing
# yy = squeeze2d(yy, 2)
scale = gain_model_params(iso)
shift = 0.0
y = x
if scale is not None:
y *= scale
if shift is not None:
y += shift
return y
def _inverse_and_log_det_jacobian(self,
y,
yy,
nlf0=None,
nlf1=None,
iso=None,
cam=None):
scale = gain_model_params(iso)
tf.summary.histogram('gain/scale', scale)
x = y
if scale is not None:
x /= scale
if scale is None:
log_abs_det_J_inv = tf.constant(0., dtype=y.dtype, name="ildj")
else:
log_abs_det_J_inv = -tf.log(scale)
if self._last_layer:
return tf.layers.flatten(x), log_abs_det_J_inv
return x, log_abs_det_J_inv