def __renderLoss(self, tile_size, batchSize, targets, outputs):
surfaceArray = helpers.generateSurfaceArray(tile_size)
rendererImpl = renderer.GGXRenderer(includeDiffuse=self.includeDiffuse)
targetsRenderings, outputsRenderings = self.__generateRenderings(
rendererImpl, batchSize, targets, outputs, surfaceArray)
reshapedTargetsRendering = tf.reshape(targetsRenderings, [
-1,
int(targetsRenderings.get_shape()[2]),
int(targetsRenderings.get_shape()[3]),
int(targetsRenderings.get_shape()[4])
])
reshapedOutputsRendering = tf.reshape(outputsRenderings, [
-1,
int(outputsRenderings.get_shape()[2]),
int(outputsRenderings.get_shape()[3]),
int(outputsRenderings.get_shape()[4])
])
currentLoss = l1(tf.log(reshapedTargetsRendering + epsilonRender),
tf.log(reshapedOutputsRendering + epsilonRender))
ssimLoss = SSIMLoss(tf.log(reshapedTargetsRendering + epsilonRender),
tf.log(reshapedOutputsRendering + epsilonRender),
1.0)
lossTotal = currentLoss + ssimLoss
return lossTotal, targetsRenderings, outputsRenderings
def __renderLoss(self):
#Generate the grid of position between -1,1 used for rendering
self.surfaceArray = helpers.generateSurfaceArray(self.crop_size)
#Initialize the renderer
rendererImpl = renderer.GGXRenderer(includeDiffuse = self.includeDiffuse)
self.targetsRenderings, self.outputsRenderings = self.__generateRenderings(rendererImpl)
#Compute the L1 loss between the renderings, epsilon are here to avoid log(0)
targetsLogged = tf.log(self.targetsRenderings + epsilonRender) # Bias could be improved to 0.1 if the network gets upset.
outputsLogged = tf.log(self.outputsRenderings + epsilonRender)
lossTotal = l1(targetsLogged, outputsLogged)#0.5 * l1(targetsLogged, outputsLogged) + l1(DX(targetsLogged), DX(outputsLogged)) + l1(DY(targetsLogged), DY(outputsLogged))
#Return the loss
return lossTotal
def __renderInputs(self, materials, renderingScene, jitterLightPos, jitterViewPos, mixMaterials):
fullSizeMixedMaterial = materials
if mixMaterials:
alpha = tf.random_uniform([1], minval=0.1, maxval=0.9, dtype=tf.float32, name="mixAlpha")
materials1 = materials[::2]
materials2 = materials[1::2]
fullSizeMixedMaterial = helpers.mixMaterials(materials1, materials2, alpha)
if self.inputImageSize >= self.tileSize :
if self.fixCrop:
xyCropping = (self.inputImageSize - self.tileSize) // 2
xyCropping = [xyCropping, xyCropping]
else:
xyCropping = tf.random_uniform([2], 0, self.inputImageSize - self.tileSize, dtype=tf.int32)
cropped_mixedMaterial = fullSizeMixedMaterial[:,:, xyCropping[0] : xyCropping[0] + self.tileSize, xyCropping[1] : xyCropping[1] + self.tileSize, :]
elif self.inputImageSize < self.tileSize:
raise Exception("Size of the input is inferior to the size of the rendering, please provide higher resolution maps")
cropped_mixedMaterial.set_shape([None, self.nbTargetsToRead, self.tileSize, self.tileSize, 3])
mixedMaterial = helpers.adaptRougness(cropped_mixedMaterial)
targetstoRender = helpers.target_reshape(mixedMaterial) #reshape it to be compatible with the rendering algorithm [?, size, size, 12]
nbRenderings = 1
rendererInstance = renderer.GGXRenderer(includeDiffuse = True)
## Do renderings of the mixedMaterial
targetstoRender = helpers.preprocess(targetstoRender) #Put targets to -1; 1
surfaceArray = helpers.generateSurfaceArray(self.tileSize)
inputs = helpers.generateInputRenderings(rendererInstance, targetstoRender, self.batchSize, nbRenderings, surfaceArray, renderingScene, jitterLightPos, jitterViewPos, self.useAmbientLight, useAugmentationInRenderings = self.useAugmentationInRenderings)
self.gammaCorrectedInputsBatch = tf.squeeze(inputs, [1])
inputs = tf.pow(inputs, 2.2) # correct gamma
if self.logInput:
inputs = helpers.logTensor(inputs)
inputs = helpers.preprocess(inputs) #Put inputs to -1; 1
targets = helpers.target_deshape(targetstoRender, self.nbTargetsToRead)
return targets, inputs
def __renderInputs(self, materials, renderingScene, jitterLightPos,
jitterViewPos, mixMaterials, isTest, renderSize):
mixedMaterial = materials
if mixMaterials:
alpha = tf.random_uniform([1],
minval=0.1,
maxval=0.9,
dtype=tf.float32,
name="mixAlpha")
#print("mat2: " + str(materials2))
materials1 = materials[::2]
materials2 = materials[1::2]
mixedMaterial = helpers.mixMaterials(materials1, materials2, alpha)
mixedMaterial.set_shape(
[None, self.nbTargetsToRead, renderSize, renderSize, 3])
mixedMaterial = helpers.adaptRougness(mixedMaterial)
#These 3 lines below tries to scale the albedos to get more variety and to randomly flatten the normals to disambiguate the normals and albedos. We did not see strong effect for these.
#if not isTest and self.useAugmentationInRenderings:
# mixedMaterial = helpers.adaptAlbedos(mixedMaterial, self.batchSize)
# mixedMaterial = helpers.adaptNormals(mixedMaterial, self.batchSize)
reshaped_targets_batch = helpers.target_reshape(
mixedMaterial
) #reshape it to be compatible with the rendering algorithm [?, size, size, 12]
nbRenderings = self.maxInputToRead
if not self.fixImageNb:
#If we don't want a constant number of input images, we randomly select a number of input images between 1 and the maximum number of images defined by the user.
nbRenderings = tf.random_uniform([1],
1,
self.maxInputToRead + 1,
dtype=tf.int32)[0]
rendererInstance = renderer.GGXRenderer(includeDiffuse=True)
## Do renderings of the mixedMaterial
targetstoRender = reshaped_targets_batch
pixelsToAdd = 0
targetstoRender = helpers.preprocess(
targetstoRender) #Put targets to -1; 1
surfaceArray = helpers.generateSurfaceArray(
renderSize, pixelsToAdd
) #Generate a grid Y,X between -1;1 to act as the pixel support of the rendering (computer the direction vector between each pixel and the light/view)
#Do the renderings
inputs = helpers.generateInputRenderings(
rendererInstance,
targetstoRender,
self.batchSize,
nbRenderings,
surfaceArray,
renderingScene,
jitterLightPos,
jitterViewPos,
self.useAmbientLight,
useAugmentationInRenderings=self.useAugmentationInRenderings)
#inputs = [helpers.preprocess(input) for input in inputs]
randomTopLeftCrop = tf.zeros([self.batchSize, nbRenderings, 2],
dtype=tf.int32)
averageCrop = 0.0
#If we want to jitter the renderings around (to try to take into account small non alignment), we should handle the material crop a bit differently
#We didn't really manage to get satisfying results with the jittering of renderings. But the code could be useful if this is of interest to Ansys.
if self.jitterRenderings:
randomTopLeftCrop = tf.random_normal(
[self.batchSize, nbRenderings, 2], 0.0,
1.0) #renderSize - self.cropSize, dtype=tf.int32)
randomTopLeftCrop = randomTopLeftCrop * tf.exp(
tf.random_normal(
[self.batchSize], 0.0,
1.0)) #renderSize - self.cropSize, dtype=tf.int32)
randomTopLeftCrop = randomTopLeftCrop - tf.reduce_mean(
randomTopLeftCrop, axis=1, keep_dims=True)
randomTopLeftCrop = tf.round(randomTopLeftCrop)
randomTopLeftCrop = tf.cast(randomTopLeftCrop, dtype=tf.int32)
averageCrop = tf.cast(self.maxJitteringPixels * 0.5,
dtype=tf.int32)
randomTopLeftCrop = randomTopLeftCrop + averageCrop
randomTopLeftCrop = tf.clip_by_value(randomTopLeftCrop, 0,
self.maxJitteringPixels)
totalCropSize = self.cropSize
inputs, targets = helpers.cutSidesOut(inputs, targetstoRender,
randomTopLeftCrop, totalCropSize,
self.firstAsGuide, averageCrop)
print("inputs shape after" + str(inputs.get_shape()))
self.gammaCorrectedInputsBatch = inputs
tf.summary.image("GammadInputs",
helpers.convert(inputs[0, :]),
max_outputs=5)
inputs = tf.pow(inputs, 2.2) # correct gamma
if self.logInput:
inputs = helpers.logTensor(inputs)
inputs = helpers.preprocess(inputs)
targets = helpers.target_deshape(targets, self.nbTargetsToRead)
return targets, inputs
def populateInNetworkFeedGraphSpatialMix(self,
renderingScene,
shuffle=True,
imageSize=512,
useSpatialMix=True):
with tf.name_scope("load_images"):
#Create a tensor out of the list of paths
filenamesTensor = tf.constant(self.pathList)
#Reads a slice of the tensor, for example, if the tensor is of shape [100,2], the slice shape should be [2] (to check if we have problem here)
dataset = tf.data.Dataset.from_tensor_slices(filenamesTensor)
#for each slice apply the __readImages function
dataset = dataset.map(self.__readImagesGT,
num_parallel_calls=int(
multiprocessing.cpu_count() / 4))
#Authorize repetition of the dataset when one epoch is over.
#shuffle = True
if shuffle:
dataset = dataset.shuffle(buffer_size=16,
reshuffle_each_iteration=True)
#set batch size
dataset = dataset.repeat()
toPull = self.batchSize
if useSpatialMix:
toPull = self.batchSize * 2
batched_dataset = dataset.batch(toPull)
batched_dataset = batched_dataset.prefetch(buffer_size=4)
#Create an iterator to be initialized
iterator = batched_dataset.make_initializable_iterator()
#Create the node to retrieve next batch
paths_batch, targets_batch = iterator.get_next()
inputRealSize = imageSize #Should be input image size but changed tmp
if useSpatialMix:
threshold = 0.5
perlinNoise = tf.expand_dims(tf.expand_dims(
helpers.generate_perlin_noise_2d(
(inputRealSize, inputRealSize), (1, 1)),
axis=-1),
axis=0)
perlinNoise = (perlinNoise + 1.0) * 0.5
perlinNoise = perlinNoise >= threshold
perlinNoise = tf.cast(perlinNoise, tf.float32)
inverted = 1.0 - perlinNoise
materialsMixed1 = targets_batch[::2] * perlinNoise
materialsMixed2 = targets_batch[1::2] * inverted
fullSizeMixedMaterial = materialsMixed1 + materialsMixed2
targets_batch = fullSizeMixedMaterial
paths_batch = paths_batch[::2]
targetstoRender = helpers.target_reshape(
targets_batch
) #reshape it to be compatible with the rendering algorithm [?, size, size, 12]
nbRenderings = 1
rendererInstance = renderer.GGXRenderer(includeDiffuse=True)
## Do renderings of the mixedMaterial
mixedMaterial = helpers.adaptRougness(targetstoRender)
targetstoRender = helpers.preprocess(
targetstoRender) #Put targets to -1; 1
surfaceArray = helpers.generateSurfaceArray(inputRealSize)
inputs_batch = helpers.generateInputRenderings(
rendererInstance,
targetstoRender,
self.batchSize,
nbRenderings,
surfaceArray,
renderingScene,
False,
False,
self.useAmbientLight,
useAugmentationInRenderings=self.useAugmentationInRenderings)
targets_batch = helpers.target_deshape(targetstoRender,
self.nbTargetsToRead)
self.gammaCorrectedInputsBatch = tf.squeeze(inputs_batch, [1])
#tf.summary.image("GammadInputs", helpers.convert(inputs[0, :]), max_outputs=5)
inputs_batch = tf.pow(inputs_batch, 2.2) # correct gamma
if self.logInput:
inputs_batch = helpers.logTensor(inputs_batch)
#Do the random crop, if the crop if fix, crop in the middle
if inputRealSize > self.tileSize:
if self.fixCrop:
xyCropping = (inputRealSize - self.tileSize) // 2
xyCropping = [xyCropping, xyCropping]
else:
xyCropping = tf.random_uniform([1],
0,
inputRealSize -
self.tileSize,
dtype=tf.int32)
inputs_batch = inputs_batch[:, :, xyCropping[0]:xyCropping[0] +
self.tileSize,
xyCropping[0]:xyCropping[0] +
self.tileSize, :]
targets_batch = targets_batch[:, :,
xyCropping[0]:xyCropping[0] +
self.tileSize,
xyCropping[0]:xyCropping[0] +
self.tileSize, :]
#Set shapes
inputs_batch = tf.squeeze(
inputs_batch, [1]
) #Before this the input has a useless dimension in 1 as we have only 1 rendering
inputs_batch.set_shape([None, self.tileSize, self.tileSize, 3])
targets_batch.set_shape(
[None, self.nbTargetsToRead, self.tileSize, self.tileSize, 3])
#Populate the object
self.stepsPerEpoch = int(
math.floor(len(self.pathList) / self.batchSize))
self.inputBatch = inputs_batch
self.targetBatch = targets_batch
self.iterator = iterator
self.pathBatch = paths_batch