def main():
if a.seed is None:
a.seed = random.randint(0, 2**31 - 1)
tf.set_random_seed(a.seed)
np.random.seed(a.seed)
random.seed(a.seed)
loadCheckpointOption(a.mode, a.checkpoint) #loads so that I don't mix up options and it generates data corresponding to this training
config = tf.ConfigProto()
if not os.path.exists(a.output_dir):
os.makedirs(a.output_dir)
with open(os.path.join(a.output_dir, "options.json"), "w") as f:
f.write(json.dumps(vars(a), sort_keys=True, indent=4))
data = dataReader.dataset(a.input_dir, imageFormat = a.imageFormat, trainFolder = a.trainFolder, testFolder = a.testFolder, nbTargetsToRead = a.nbTargets, tileSize=TILE_SIZE, inputImageSize=a.input_size, batchSize=a.batch_size, fixCrop = (a.mode == "test"), mixMaterials = (a.mode == "train" or a.mode == "finetune"), logInput = a.useLog, useAmbientLight = a.useAmbientLight, useAugmentationInRenderings = not a.NoAugmentationInRenderings)
# Populate data
data.loadPathList(a.inputMode, a.mode, a.mode == "train" or a.mode == "finetune", inputpythonList)
if a.feedMethod == "render":
if a.mode == "train":
data.populateInNetworkFeedGraph(a.renderingScene, a.jitterLightPos, a.jitterViewPos, shuffle = (a.mode == "train" or a.mode == "finetune"))
elif a.mode == "finetune":
data.populateInNetworkFeedGraphSpatialMix(a.renderingScene, shuffle = False, imageSize = a.input_size)
elif a.feedMethod == "files":
data.populateFeedGraph(shuffle = (a.mode == "train" or a.mode == "finetune"))
if a.mode == "train" or a.mode == "finetune":
with tf.name_scope("recurrentTest"):
dataTest = dataReader.dataset(a.input_dir, imageFormat = a.imageFormat, testFolder = a.testFolder, nbTargetsToRead = a.nbTargets, tileSize=TILE_SIZE, inputImageSize=a.test_input_size, batchSize=a.batch_size, fixCrop = True, mixMaterials = False, logInput = a.useLog, useAmbientLight = a.useAmbientLight, useAugmentationInRenderings = not a.NoAugmentationInRenderings)
dataTest.loadPathList(a.inputMode, "test", False, inputpythonList)
if a.testApproach == "render":
#dataTest.populateInNetworkFeedGraphSpatialMix(a.renderingScene, shuffle = False, imageSize = TILE_SIZE, useSpatialMix=False)
dataTest.populateInNetworkFeedGraph(a.renderingScene, a.jitterLightPos, a.jitterViewPos, shuffle = False)
elif a.testApproach == "files":
dataTest.populateFeedGraph(False)
targetsReshaped = helpers.target_reshape(data.targetBatch)
#CreateModel
model = mod.Model(data.inputBatch, generatorOutputChannels=9)
model.create_model()
if a.mode == "train" or a.mode == "finetune":
testTargetsReshaped = helpers.target_reshape(dataTest.targetBatch)
testmodel = mod.Model(dataTest.inputBatch, generatorOutputChannels=9, reuse_bool=True)
testmodel.create_model()
display_fetches_test, _ = helpers.display_images_fetches(dataTest.pathBatch, dataTest.inputBatch, dataTest.targetBatch, dataTest.gammaCorrectedInputsBatch, testmodel.output, a.nbTargets, a.logOutputAlbedos)
loss = losses.Loss(a.loss, model.output, targetsReshaped, TILE_SIZE, a.batch_size, tf.placeholder(tf.float64, shape=(), name="lr"), a.includeDiffuse, a.nbSpecularRendering, a.nbDiffuseRendering)
loss.createLossGraph()
loss.createTrainVariablesGraph()
#Register Renderings And Loss In Tensorflow
display_fetches, converted_images = helpers.display_images_fetches(data.pathBatch, data.inputBatch, data.targetBatch, data.gammaCorrectedInputsBatch, model.output, a.nbTargets, a.logOutputAlbedos)
if a.mode == "train":
helpers.registerTensorboard(data.pathBatch, converted_images, a.nbTargets, loss.lossValue, a.batch_size, loss.targetsRenderings, loss.outputsRenderings)
#Run either training or test
with tf.name_scope("parameter_count"):
parameter_count = tf.reduce_sum([tf.reduce_prod(tf.shape(v)) for v in tf.trainable_variables()])
saver = tf.train.Saver(max_to_keep=1)
if a.checkpoint is not None:
print("reading model from checkpoint : " + a.checkpoint)
checkpoint = tf.train.latest_checkpoint(a.checkpoint)
partialSaver = helpers.optimistic_saver(checkpoint) #Be careful this will silently not load variables if they are missing from the graph or checkpoint
logdir = a.output_dir if a.summary_freq > 0 else None
sv = tf.train.Supervisor(logdir=logdir, save_summaries_secs=0, saver=None)
with sv.managed_session("", config= config) as sess:
sess.run(data.iterator.initializer)
print("parameter_count =", sess.run(parameter_count))
if a.checkpoint is not None:
print("restoring model from checkpoint : " + a.checkpoint)
partialSaver.restore(sess, checkpoint)
max_steps = 2**32
if a.max_epochs is not None:
max_steps = data.stepsPerEpoch * a.max_epochs
if a.max_steps is not None:
max_steps = a.max_steps
sess.run(data.iterator.initializer)
if a.mode == "test":
filesets = test(sess, data, max_steps, display_fetches, output_dir = a.output_dir)
if a.mode == "train" or a.mode == "finetune":
train(sv, sess, data, max_steps, display_fetches, display_fetches_test, dataTest, saver, loss, a.output_dir)
def main():
if a.seed is None:
a.seed = random.randint(0, 2**31 - 1)
tf.set_random_seed(a.seed)
np.random.seed(a.seed)
random.seed(a.seed)
#Load some options from the checkpoint if we provided one.
loadCheckpointOption()
#If we feed the network with renderings done in the network for a test run, we save the images before, to be able to compare later with other networks on the same testset.
if a.mode == "test" and a.feedMethod == "render":
testHelpers.renderTests(a.input_dir, a.testFolder, a.maxImages,
tmpFolder, a.imageFormat, CROP_SIZE,
a.nbTargets, a.input_size, a.batch_size,
a.renderingScene, a.jitterLightPos,
a.jitterViewPos, a.inputMode, a.mode,
a.output_dir)
generateTmpData = True
a.nbInputs = a.maxImages
a.feedMethod = "files"
a.testFolder = tmpFolder
a.input_size = CROP_SIZE
backupOutputDir = a.output_dir
#We run the network once if we a training
nbRun = 1
#And as many time as the maximum number of images we want to treat with if testing (to have results with one image, two images, three images etc... to see the improvement)
if a.mode == "test":
nbRun = a.maxImages #1
a.fixImageNb = True
#Now run the network nbRun times.
for runID in range(nbRun):
maxInputNb = a.maxImages
if a.mode == "test":
maxInputNb = runID + 1 #a.maxImages
a.output_dir = os.path.join(backupOutputDir, str(runID))
tf.reset_default_graph()
#Create the output dir if it doesn't exist
if not os.path.exists(a.output_dir):
os.makedirs(a.output_dir)
#Write to the "options" file the different parameters of this run.
with open(os.path.join(a.output_dir, "options.json"), "w") as f:
f.write(json.dumps(vars(a), sort_keys=True, indent=4))
#Create a dataset object
data = dataReader.dataset(
a.input_dir,
imageFormat=a.imageFormat,
trainFolder=a.trainFolder,
testFolder=a.testFolder,
inputNumbers=a.nbInputs,
maxInputToRead=maxInputNb,
nbTargetsToRead=a.nbTargets,
cropSize=CROP_SIZE,
inputImageSize=a.input_size,
batchSize=a.batch_size,
fixCrop=(a.mode == "test"),
mixMaterials=(a.mode == "train"),
fixImageNb=a.fixImageNb,
logInput=a.useLog,
useAmbientLight=a.useAmbientLight,
jitterRenderings=a.jitterRenderings,
firstAsGuide=False,
useAugmentationInRenderings=not a.NoAugmentationInRenderings,
mode=a.mode)
# Populate the list of files the dataset will contain
data.loadPathList(a.inputMode, a.mode, a.mode == "train")
# Depending on wheter we want to render our input data or directly use files, we create the tensorflow data loading system.
if a.feedMethod == "render":
data.populateInNetworkFeedGraph(a.renderingScene,
a.jitterLightPos,
a.jitterViewPos,
a.mode == "test",
shuffle=a.mode == "train")
elif a.feedMethod == "files":
data.populateFeedGraph(shuffle=a.mode == "train")
# Here we reshape the input to have all the images in the first dimension (to treat in parallel)
inputReshaped, dyn_batch_size = helpers.input_reshape(
data.inputBatch, a.NoMaxPooling, a.maxImages)
if a.mode == "train":
with tf.name_scope("recurrentTest"):
#Initialize different data for tests.
dataTest = dataReader.dataset(
a.input_dir,
imageFormat=a.imageFormat,
testFolder=a.testFolder,
inputNumbers=a.nbInputs,
maxInputToRead=a.maxImages,
nbTargetsToRead=a.nbTargets,
cropSize=CROP_SIZE,
inputImageSize=a.input_size,
batchSize=a.batch_size,
fixCrop=True,
mixMaterials=False,
fixImageNb=a.fixImageNb,
logInput=a.useLog,
useAmbientLight=a.useAmbientLight,
jitterRenderings=a.jitterRenderings,
firstAsGuide=a.firstAsGuide,
useAugmentationInRenderings=not a.
NoAugmentationInRenderings,
mode=a.mode)
dataTest.loadPathList(a.inputMode, "test", False)
if a.feedMethod == "render":
dataTest.populateInNetworkFeedGraph(a.renderingScene,
a.jitterLightPos,
a.jitterViewPos,
True,
shuffle=False)
elif a.feedMethod == "files":
dataTest.populateFeedGraph(False)
TestinputReshaped, test_dyn_batch_size = helpers.input_reshape(
dataTest.inputBatch, a.NoMaxPooling, a.maxImages)
#Reshape the targets to [?(Batchsize), 256,256,12]
targetsReshaped = helpers.target_reshape(data.targetBatch)
#Create the object to contain the network model.
model = mod.Model(inputReshaped,
dyn_batch_size,
last_convolutions_channels=last_convs_chans,
generatorOutputChannels=64,
useCoordConv=a.useCoordConv,
firstAsGuide=a.firstAsGuide,
NoMaxPooling=a.NoMaxPooling,
pooling_type=a.poolingtype)
#Initialize the model.
model.create_model()
if a.mode == "train":
#Initialize the regular test network with different data so that it can run regular test sets.
testTargetsReshaped = helpers.target_reshape(dataTest.targetBatch)
testmodel = mod.Model(TestinputReshaped,
test_dyn_batch_size,
last_convolutions_channels=last_convs_chans,
generatorOutputChannels=64,
reuse_bool=True,
useCoordConv=a.useCoordConv,
firstAsGuide=a.firstAsGuide,
NoMaxPooling=a.NoMaxPooling,
pooling_type=a.poolingtype)
testmodel.create_model()
#Organize the images we want to retrieve from the test network run
display_fetches_test, _ = helpers.display_images_fetches(
dataTest.pathBatch, dataTest.inputBatch, dataTest.targetBatch,
dataTest.gammaCorrectedInputsBatch, testmodel.output,
a.nbTargets, a.logOutputAlbedos)
# Compute the training network loss.
loss = losses.Loss(a.loss, model.output, targetsReshaped,
CROP_SIZE, a.batch_size,
tf.placeholder(tf.float64, shape=(),
name="lr"), a.includeDiffuse)
loss.createLossGraph()
#Create the training graph part
loss.createTrainVariablesGraph()
#Organize the images we want to retrieve from the train network run
display_fetches, converted_images = helpers.display_images_fetches(
data.pathBatch, data.inputBatch, data.targetBatch,
data.gammaCorrectedInputsBatch, model.output, a.nbTargets,
a.logOutputAlbedos)
if a.mode == "train":
#Register inputs, targets, renderings and loss in Tensorboard
helpers.registerTensorboard(data.pathBatch, converted_images,
a.maxImages, a.nbTargets,
loss.lossValue, a.batch_size,
loss.targetsRenderings,
loss.outputsRenderings)
#Compute how many paramters the network has
with tf.name_scope("parameter_count"):
parameter_count = tf.reduce_sum([
tf.reduce_prod(tf.shape(v)) for v in tf.trainable_variables()
])
#Initialize a saver
saver = tf.train.Saver(max_to_keep=1)
if a.checkpoint is not None:
print("reading model from checkpoint : " + a.checkpoint)
checkpoint = tf.train.latest_checkpoint(a.checkpoint)
partialSaver = helpers.optimistic_saver(checkpoint)
logdir = a.output_dir if a.summary_freq > 0 else None
sv = tf.train.Supervisor(logdir=logdir,
save_summaries_secs=0,
saver=None)
#helpers.print_trainable()
with sv.managed_session() as sess:
print("parameter_count =", sess.run(parameter_count))
#Loads the checkpoint
if a.checkpoint is not None:
print("restoring model from checkpoint : " + a.checkpoint)
partialSaver.restore(sess, checkpoint)
#Evaluate how many steps to run
max_steps = 2**32
if a.max_epochs is not None:
max_steps = data.stepsPerEpoch * a.max_epochs
if a.max_steps is not None:
max_steps = a.max_steps
#If we want to run a test
if a.mode == "test" or a.mode == "eval":
filesets = test(sess,
data,
max_steps,
display_fetches,
output_dir=a.output_dir)
if runID == nbRun - 1 and runID >= 1: #If we are at the last iteration of the test, generate the full html
helpers.writeGlobalHTML(backupOutputDir, filesets,
a.nbTargets, a.mode, a.maxImages)
#If we want to train
if a.mode == "train":
train(sv, sess, data, max_steps, display_fetches,
display_fetches_test, dataTest, saver, loss)
def renderTests(input_dir, testFolder, maxInputNb, tmpFolder, imageFormat,
CROP_SIZE, nbTargets, input_size, batchSize, renderingScene,
jitterLightPos, jitterViewPos, inputMode, mode, outputDir):
fullOutputDir = os.path.join(outputDir, "testGenerationLog")
if not os.path.exists(fullOutputDir):
os.makedirs(fullOutputDir)
fullTmpFolder = os.path.join(input_dir, tmpFolder)
if os.path.exists(fullTmpFolder):
shutil.rmtree(fullTmpFolder)
os.makedirs(fullTmpFolder)
data = dataReader.dataset(input_dir,
imageType=imageFormat,
testFolder=testFolder,
maxInputToRead=maxInputNb,
nbTargetsToRead=nbTargets,
cropSize=CROP_SIZE,
inputImageSize=input_size,
batchSize=batchSize,
fixCrop=True,
mixMaterials=True,
fixImageNb=True,
logInput=False,
useAmbientLight=False,
jitterRenderings=False,
useAugmentationInRenderings=False)
data.loadPathList(inputMode, mode, False)
data.maxJitteringPixels = 0 #if 0 here, will produce pixel perfect renderings
data.populateInNetworkFeedGraph(renderingScene,
jitterLightPos,
jitterViewPos,
True,
shuffle=True)
data.gammaCorrectedInputsBatch.set_shape(
[batchSize, maxInputNb, None, None, None])
display_fetches = display_images_fetches_fullTest(
data.pathBatch, data.gammaCorrectedInputsBatch, data.targetBatch,
nbTargets
) #display_images_fetches(data.pathBatch, data.gammaCorrectedInputsBatch, data.targetBatch, nbTargets) # save the gamma corrected version of the inputs
sv = tf.train.Supervisor(logdir=fullOutputDir,
save_summaries_secs=0,
saver=None)
with sv.managed_session() as sess:
max_steps = 2**32
sess.run(data.iterator.initializer)
print(data.stepsPerEpoch)
max_steps = 100 #data.stepsPerEpoch
for step in range(max_steps):
try:
results = sess.run(display_fetches)
save_images_fullPath(
results, fullTmpFolder, batchSize, nbTargets, step
) #save_images(results, fullTmpFolder, batchSize, nbTargets)
except tf.errors.OutOfRangeError:
print("testing fails in OutOfRangeError")
continue
print("RENDERINGS DONE")