def __l1Loss(self):
#outputs have shape [?, height, width, 12]
#targets have shape [?, height, width, 12]
outputsNormal = self.outputs[:,:,:,0:3]
outputsDiffuse = tf.log(epsilonL1 + helpers.deprocess(self.outputs[:,:,:,3:6]))
outputsRoughness = self.outputs[:,:,:,6:9]
outputsSpecular = tf.log(epsilonL1 + helpers.deprocess(self.outputs[:,:,:,9:12]))
targetsNormal = self.targets[:,:,:,0:3]
targetsDiffuse = tf.log(epsilonL1 + helpers.deprocess(self.targets[:,:,:,3:6]))
targetsRoughness = self.targets[:,:,:,6:9]
targetsSpecular = tf.log(epsilonL1 + helpers.deprocess(self.targets[:,:,:,9:12]))
return l1(outputsNormal, targetsNormal) + l1(outputsDiffuse, targetsDiffuse) + l1(outputsRoughness, targetsRoughness) + l1(outputsSpecular, targetsSpecular)
def deprocess_images_fullTest(inputs, targets, nbTargets):
targets = helpers.deprocess(targets)
with tf.name_scope("transform_images"):
targetShape = targets.get_shape()
targets_reshaped = concat_tensor_display(targets,
axisToConcat=2,
axisToSplit=1)
tensorOneInput, SurfaceLightFixedView, HemishpereLightFixedView = concatSplitInputs(
inputs, axisToConcat=2,
axisToSplit=1) #HemishpereLightHemisphereView
tensorOneInput = tf.concat([tensorOneInput, targets_reshaped], axis=2)
SurfaceLightFixedView = tf.concat(
[SurfaceLightFixedView, targets_reshaped], axis=2)
HemishpereLightFixedView = tf.concat(
[HemishpereLightFixedView, targets_reshaped], axis=2)
#HemishpereLightHemisphereView = tf.concat([HemishpereLightHemisphereView, targets_reshaped], axis = 2)
with tf.name_scope("convert_images"):
tensorOneInput = helpers.convert(tensorOneInput)
SurfaceLightFixedView = helpers.convert(SurfaceLightFixedView)
HemishpereLightFixedView = helpers.convert(HemishpereLightFixedView)
#HemishpereLightHemisphereView = helpers.convert(HemishpereLightHemisphereView)
return tensorOneInput, SurfaceLightFixedView, HemishpereLightFixedView #, HemishpereLightHemisphereView
def tf_calculateBRDF(self, svbrdf, wiNorm, woNorm, currentConeTargetPos,
currentLightPos, multiLight):
h = helpers.tf_Normalize(tf.add(wiNorm, woNorm) / 2.0)
#Put all the parameter values between 0 and 1 except the normal as they should be used between -1 and 1 (as they express a direction in a 360° sphere)
diffuse = tf.expand_dims(helpers.squeezeValues(
helpers.deprocess(svbrdf[:, :, :, 3:6]), 0.0, 1.0),
axis=1)
normals = tf.expand_dims(svbrdf[:, :, :, 0:3], axis=1)
normals = helpers.tf_Normalize(normals)
specular = tf.expand_dims(helpers.squeezeValues(
helpers.deprocess(svbrdf[:, :, :, 9:12]), 0.0, 1.0),
axis=1)
roughness = tf.expand_dims(helpers.squeezeValues(
helpers.deprocess(svbrdf[:, :, :, 6:9]), 0.0, 1.0),
axis=1)
#Avoid roughness = 0 to avoid division by 0
roughness = tf.maximum(roughness, 0.001)
#If we have multiple lights to render, add a dimension to handle it.
if multiLight:
diffuse = tf.expand_dims(diffuse, axis=1)
normals = tf.expand_dims(normals, axis=1)
specular = tf.expand_dims(specular, axis=1)
roughness = tf.expand_dims(roughness, axis=1)
NdotH = helpers.tf_DotProduct(normals, h)
NdotL = helpers.tf_DotProduct(normals, wiNorm)
NdotV = helpers.tf_DotProduct(normals, woNorm)
VdotH = helpers.tf_DotProduct(woNorm, h)
diffuse_rendered = self.tf_diffuse(diffuse, specular)
D_rendered = self.tf_D(roughness, tf.maximum(0.0, NdotH))
G_rendered = self.tf_G(roughness, tf.maximum(0.0, NdotL),
tf.maximum(0.0, NdotV))
F_rendered = self.tf_F(specular, tf.maximum(0.0, VdotH))
specular_rendered = F_rendered * (G_rendered * D_rendered * 0.25)
result = specular_rendered
#Add the diffuse part of the rendering if required.
if self.includeDiffuse:
result = result + diffuse_rendered
return result, NdotL
def tf_calculateBRDF(self, svbrdf, wiNorm, woNorm, currentConeTargetPos,
currentLightPos, multiLight, lossRendering):
h = helpers.tf_Normalize(tf.add(wiNorm, woNorm) / 2.0)
diffuse = tf.expand_dims(helpers.squeezeValues(
helpers.deprocess(svbrdf[:, :, :, 3:6]), 0.0, 1.0),
axis=1)
normals = tf.expand_dims(svbrdf[:, :, :, 0:3], axis=1)
normals = helpers.tf_Normalize(normals)
specular = tf.expand_dims(helpers.squeezeValues(
helpers.deprocess(svbrdf[:, :, :, 9:12]), 0.0, 1.0),
axis=1)
roughness = tf.expand_dims(helpers.squeezeValues(
helpers.deprocess(svbrdf[:, :, :, 6:9]), 0.0, 1.0),
axis=1)
roughness = tf.maximum(roughness, 0.001)
#This is the simulation of ambient lighting for fine tuning mostly
if not lossRendering:
diffuse = diffuse + (0.15 * specular)
if multiLight:
diffuse = tf.expand_dims(diffuse, axis=1)
normals = tf.expand_dims(normals, axis=1)
specular = tf.expand_dims(specular, axis=1)
roughness = tf.expand_dims(roughness, axis=1)
NdotH = helpers.tf_DotProduct(normals, h)
NdotL = helpers.tf_DotProduct(normals, wiNorm)
NdotV = helpers.tf_DotProduct(normals, woNorm)
VdotH = helpers.tf_DotProduct(woNorm, h)
diffuse_rendered = self.tf_diffuse(diffuse, specular)
D_rendered = self.tf_D(roughness, tf.maximum(0.0, NdotH))
G_rendered = self.tf_G(roughness, tf.maximum(0.0, NdotL),
tf.maximum(0.0, NdotV))
F_rendered = self.tf_F(specular, tf.maximum(0.0, VdotH))
specular_rendered = F_rendered * (G_rendered * D_rendered * 0.25)
result = specular_rendered
if self.includeDiffuse:
result = result + diffuse_rendered
return result, NdotL
def deprocess_images(inputs, targets, nbTargets):
#inputs = helpers.deprocess(inputs)
targets = helpers.deprocess(targets)
with tf.name_scope("transform_images"):
targetShape = targets.get_shape()
targets_reshaped = concat_tensor_display(targets,
axisToConcat=2,
axisToSplit=1)
inputs_reshaped = tf.reshape(inputs, [
-1,
int(inputs.get_shape()[2]),
int(inputs.get_shape()[3]),
int(inputs.get_shape()[4])
])
tensorToSave = tf.concat([inputs_reshaped, targets_reshaped], axis=2)
with tf.name_scope("convert_images"):
tensorToSave = helpers.convert(tensorToSave)
return tensorToSave