def generate_tf_record(data_dir,
raw_data=False,
tfrecord_path="serialized_dataset",
num_shards=8):
teacher_sett = settings.Settings(use_student_settings=False)
student_sett = settings.Settings(use_student_settings=True)
dataset_args = teacher_sett["dataset"]
if dataset_args["name"].lower().strip() == "div2k":
assert len(data_dir) == 2
ds = dataset.load_div2k_dataset(data_dir[0],
data_dir[1],
student_sett["hr_size"],
shuffle=True)
elif raw_data:
ds = dataset.load_dataset_directory(
dataset_args["name"], data_dir,
dataset.scale_down(method=dataset_args["scale_method"],
size=student_sett["hr_size"]))
else:
ds = dataset.load_dataset(dataset_args["name"],
dataset.scale_down(
method=dataset_args["scale_method"],
size=student_sett["hr_size"]),
data_dir=data_dir)
to_tfrecord(ds, tfrecord_path, num_shards)
def __init__(
self,
teacher,
discriminator,
summary_writer,
summary_writer_2=None,
model_dir="",
data_dir="",
strategy=None):
"""
Args:
teacher: Keras Model of pre-trained teacher generator.
(Generator of ESRGAN)
discriminator: Keras Model of pre-trained teacher discriminator.
(Discriminator of ESRGAN)
summary_writer: tf.summary.SummaryWriter object for writing
summary for Tensorboard.
data_dir: Location of the stored dataset.
raw_data: Indicate if data_dir contains Raw Data or TFRecords.
model_dir: Location to store checkpoints and SavedModel directory.
"""
self.teacher_generator = teacher
self.teacher_discriminator = discriminator
self.teacher_settings = settings.Settings(use_student_settings=False)
self.student_settings = settings.Settings(use_student_settings=True)
self.model_dir = model_dir
self.strategy = strategy
self.train_args = self.student_settings["train"]
self.batch_size = self.teacher_settings["batch_size"]
self.hr_size = self.student_settings["hr_size"]
self.lr_size = tf.unstack(self.hr_size)[:-1]
self.lr_size.append(tf.gather(self.hr_size, len(self.hr_size) - 1) * 4)
self.lr_size = tf.stack(self.lr_size) // 4
self.summary_writer = summary_writer
self.summary_writer_2 = summary_writer_2
# Loading TFRecord Dataset
self.dataset = dataset.load_dataset(
data_dir,
lr_size=self.lr_size,
hr_size=self.hr_size)
self.dataset = (
self.dataset.repeat()
.batch(self.batch_size, drop_remainder=True)
.prefetch(1024))
self.dataset = iter(self.strategy.experimental_distribute_dataset(
self.dataset))
# Reloading Checkpoint from Phase 2 Training of ESRGAN
checkpoint = tf.train.Checkpoint(
G=self.teacher_generator,
D=self.teacher_discriminator)
utils.load_checkpoint(
checkpoint,
"phase_2",
basepath=model_dir,
use_student_settings=False)
def init(self, first_call=True):
self.settings = settings.Settings(use_student_settings=True)
self._scale_factor = self.settings["scale_factor"]
self._scale_value = self.settings["scale_value"]
rrdb_student_config = self.settings["student_config"]["rrdb_student"]
rrdb_block = partial(ResidualInResidualBlock, first_call=first_call)
growth_channels = rrdb_student_config["growth_channels"]
depthwise_conv = partial(tf.keras.layers.DepthwiseConv2D,
kernel_size=[3, 3],
strides=[1, 1],
use_bias=True,
padding="same")
convolution = partial(tf.keras.layers.Conv2D,
kernel_size=[3, 3],
use_bias=True,
strides=[1, 1],
padding="same")
conv_transpose = partial(tf.keras.layers.Conv2DTranspose,
kernel_size=[3, 3],
use_bias=True,
strides=self._scale_value,
padding="same")
self._rrdb_trunk = tf.keras.Sequential(
[rrdb_block() for _ in range(rrdb_student_config["trunk_size"])])
self._first_conv = convolution(filters=64)
self._upsample_layers = {
"upsample_%d" % index: conv_transpose(filters=growth_channels)
for index in range(1, self._scale_factor)
}
key = "upsample_%d" % self._scale_factor
self._upsample_layers[key] = conv_transpose(filters=3)
self._conv_last = depthwise_conv()
self._lrelu = tf.keras.layers.LeakyReLU(alpha=0.2)
def init(self):
sett = settings.Settings(use_student_settings=True)
model_args = sett["student_config"]["vgg_student"]
self._scale_factor = sett["scale_factor"]
self._scale_value = sett["scale_value"]
depth = model_args["trunk_depth"] # Minimum 2 for scale factor of 4
depthwise_convolution = partial(tf.keras.layers.DepthwiseConv2D,
kernel_size=[3, 3],
padding="same",
use_bias=model_args["use_bias"])
conv_transpose = partial(tf.keras.layers.Conv2DTranspose,
kernel_size=[3, 3],
strides=self._scale_value,
padding="same")
convolution = partial(tf.keras.layers.Conv2D,
kernel_size=[3, 3],
strides=[1, 1],
use_bias=model_args["use_bias"],
padding="same")
trunk_depth = depth - self._scale_factor
self._conv_layers = {
"conv_%d" % index: depthwise_convolution()
for index in range(1, trunk_depth + 1)
}
self._upsample_layers = {
"upsample_%d" % index: conv_transpose(filters=32)
for index in range(1, self._scale_factor)
}
self._last_layer = conv_transpose(filters=3)
def init(self):
self.student_settings = settings.Settings(use_student_settings=True)
self._scale_factor = self.student_settings["scale_factor"]
self._scale_value = self.student_settings["scale_value"]
model_args = self.student_settings["student_config"][
"residual_student"]
depth = model_args["trunk_depth"]
depthwise_convolution = partial(tf.keras.layers.DepthwiseConv2D,
kernel_size=[3, 3],
strides=[1, 1],
padding="same",
use_bias=model_args["use_bias"])
convolution = partial(tf.keras.layers.Conv2D,
kernel_size=[3, 3],
strides=[1, 1],
padding="same",
use_bias=model_args["use_bias"])
conv_transpose = partial(tf.keras.layers.Conv2DTranspose,
kernel_size=[3, 3],
strides=self._scale_value,
padding="same")
self._lrelu = tf.keras.layers.LeakyReLU(alpha=0.2)
self._residual_scale = model_args["residual_scale_beta"]
self._conv_layers = {
"conv_%d" % index: depthwise_convolution()
for index in range(1, depth + 1)
}
self._upscale_layers = {
"upscale_%d" % index: conv_transpose(filters=32)
for index in range(1, self._scale_factor)
}
self._last_layer = conv_transpose(filters=3)
def __init__(self, first_call=True):
super(ResidualInResidualBlock, self).__init__()
self.settings = settings.Settings(use_student_settings=True)
rrdb_config = self.settings["student_config"]["rrdb_student"][
"rrdb_config"]
self._rdb_layers = {
"rdb_%d" % index: ResidualDenseBlock(first_call=first_call)
for index in range(1, rrdb_config["rdb_units"])
}
self._beta = rrdb_config["residual_scale_beta"]
def export_tflite(config="", modeldir="", mode="", **kwargs):
"""
Exports SavedModel(if not present) TFLite of the student generator.
Args:
config: Path to config file of the student.
modeldir: Path to export the SavedModel and the TFLite to.
mode: Mode of training to export. (Advsersarial / comparative)
"""
# TODO (@captain-pool): Fix Quantization and mention them in the args list.
lazy = lazy_loader.LazyLoader()
lazy.import_("teacher_imports", parent="libs", return_=False)
lazy.import_("utils", parent="libs", return_=False)
lazy.import_("dataset", parent="libs", return_=False)
globals().update(lazy.import_dict)
status = None
sett = settings.Settings(config, use_student_settings=True)
stats = settings.Stats(os.path.join(sett.path, "stats.yaml"))
student_name = sett["student_network"]
student_generator = model.Registry.models[student_name](first_call=False)
ckpt = tf.train.Checkpoint(student_generator=student_generator)
logging.info("Initiating Variables. Tracing Function.")
student_generator.predict(tf.random.normal([1, 180, 320, 3]))
if stats.get(mode):
status = utils.load_checkpoint(ckpt,
"%s_checkpoint" % mode,
basepath=modeldir,
use_student_settings=True)
if not status:
raise IOError("No checkpoint found to restore")
saved_model_dir = os.path.join(modeldir, "compressed_esrgan")
if not tf.io.gfile.exists(os.path.join(saved_model_dir, "saved_model.pb")):
tf.saved_model.save(student_generator, saved_model_dir)
converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
# TODO (@captain-pool): Try to fix Qunatization
# Current Error: Cannot Quantize LEAKY_RELU and CONV2D_TRANSPOSE
# Quantization Code Fragment
# converter.target_spec.supported_types = [tf.float16]
# converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
# converter.representative_dataset = representative_dataset_gen(
# kwargs["calibration_steps"],
# kwargs["datadir"],
# sett["hr_size"],
# [180, 320, 3])
tflite_model = converter.convert()
tflite_path = os.path.join(modeldir, "tflite", "compressed_esrgan.tflite")
with tf.io.gfile.GFile(tflite_path, "wb") as f:
f.write(tflite_model)
logging.info("Successfully writen the TFLite to: %s" % tflite_path)
def save_checkpoint(checkpoint, name, basepath="", use_student_settings=False):
""" Saves Checkpoint
Args:
checkpoint: tf.train.Checkpoint object to save.
name: name of the checkpoint to save.
basepath: base directory where checkpoint should be saved
student: boolean to indicate if settings of the student should be used.
"""
sett = settings.Settings(use_student_settings=use_student_settings)
dir_ = os.path.join(basepath, sett["checkpoint_path"][name])
logging.info("Saving checkpoint: %s Path: %s" % (name, dir_))
prefix = os.path.join(dir_, os.path.basename(dir_))
checkpoint.save(file_prefix=prefix)
def checkpoint_exists(names, basepath="", use_student_settings=False):
sett = settings.Settings(use_student_settings=use_student_settings)
if tf.nest.is_nested(names):
if isinstance(names, dict):
return False
else:
names = [names]
values = []
for name in names:
dir_ = os.path.join(basepath, sett["checkpoint_path"][name],
"checkpoint")
values.append(tf.io.gfile.exists(dir_))
return any(values)
def load_checkpoint(checkpoint, name, basepath="", use_student_settings=False):
""" Restores Checkpoint
Args:
checkpoint: tf.train.Checkpoint object to restore.
name: name of the checkpoint to restore.
basepath: base directory where checkpoint is located.
student: boolean to indicate if settings of the student should be used.
"""
sett = settings.Settings(use_student_settings=use_student_settings)
dir_ = os.path.join(basepath, sett["checkpoint_path"][name])
if tf.io.gfile.exists(os.path.join(dir_, "checkpoint")):
logging.info("Found checkpoint: %s Path: %s" % (name, dir_))
status = checkpoint.restore(tf.train.latest_checkpoint(dir_))
return status
logging.info("No Checkpoint found for %s" % name)
def __init__(self, first_call=True):
super(ResidualDenseBlock, self).__init__()
self.settings = settings.Settings(use_student_settings=True)
rdb_config = self.settings["student_config"]["rrdb_student"][
"rdb_config"]
convolution = partial(tf.keras.layers.DepthwiseConv2D,
depthwise_initializer="he_normal",
bias_initializer="he_normal",
kernel_size=[3, 3],
strides=[1, 1],
padding="same")
self._first_call = first_call
self._conv_layers = {
"conv_%d" % index: convolution()
for index in range(1, rdb_config["depth"])
}
self._lrelu = tf.keras.layers.LeakyReLU(alpha=0.2)
self._beta = rdb_config["residual_scale_beta"]
avg_min = POST.Evaluation.calc_avg_mse(root_dir)
# best avg values form middlebury evaluation page
# http://vision.middlebury.edu/stereo/eval3/
best_avg_sparse = 4.61
best_avg_dense = 12.9
data = [avg_min,best_avg_sparse,best_avg_dense]
name = 'avg_mse_bar_plot'
VISUALIZE.bar_chart(data, root_dir, name)
cm = CAMERA_MODEL.CameraModel(329.115520046, 329.115520046, 320.0, 240.0)
image_loader = IMAGE_LOADER.ImageLoader(images_dir)
vo = VO_PARSER.VoParserSynth(odom_path)
runtime_settings = SETTINGS.Settings(10, 0.0001, 0.8, 10.0, 0.05, 1.0) # TODO ADD RESCALE FACTOR HERE
ref_count = 10
ref_list = []
for count in range(1, ref_count + 1):
ref_list.append(key + count)
stereo_pair_builder = STEREO_PAIR_BUILDER.StereoPairBuilder(cm, image_loader, vo, 0, runtime_settings)
runner = RUNNER.Runner(stereo_pair_builder, runtime_settings, key, ref_list, root_dir)
ground_truth = POST.Evaluation.load_ground_truth(depth_path, 480, 640)
inverted_ground_truth = POST.Evaluation.calc_inverse_ground_truth(ground_truth, runtime_settings, 329.115520046)
# VISUALIZE.show_frame(inverted_ground_truth, runtime_settings, path=root_dir, name='ground_truth', cmap='nipy_spectral')
runner.run(VISUALIZE.visualize_enum.SHOW_DEPTH, inverted_ground_truth, normalize=True, calc_error_metrics=False,
post_process=True, regularize=True,
show_frame=True, skip_guards=False)
def train_and_export(**kwargs):
""" Train and Export Compressed ESRGAN
Args:
config: path to config file.
logdir: path to logging directory
modeldir: Path to store the checkpoints and exported model.
datadir: Path to custom data directory.
manual: Boolean to indicate if `datadir` contains Raw Files(True) / TFRecords (False)
"""
lazy = lazy_loader.LazyLoader()
student_settings = settings.Settings(kwargs["config"],
use_student_settings=True)
# Lazy importing dependencies from teacher
lazy.import_("teacher_imports", parent="libs", return_=False)
lazy.import_("teacher", parent="libs.models", return_=False)
lazy.import_("train", parent="libs", return_=False)
lazy.import_("utils", parent="libs", return_=False)
globals().update(lazy.import_dict)
tf.random.set_seed(10)
teacher_settings = settings.Settings(student_settings["teacher_config"],
use_student_settings=False)
stats = settings.Stats(os.path.join(student_settings.path, "stats.yaml"))
strategy = utils.SingleDeviceStrategy()
if kwargs["tpu"]:
cluster_resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
kwargs["tpu"])
tf.config.experimental_connect_to_host(cluster_resolver.get_master())
tf.tpu.experimental.initialize_tpu_system(cluster_resolver)
strategy = tf.distribute.experimental.TPUStrategy(cluster_resolver)
device_name = utils.assign_to_worker(kwargs["tpu"])
with tf.device(device_name), strategy.scope():
summary_writer = tf.summary.create_file_writer(
os.path.join(kwargs["logdir"], "student"))
teacher_summary_writer = tf.summary.create_file_writer(
os.path.join(kwargs["logdir"], "teacher"))
teacher_generator = teacher.generator(out_channel=3, first_call=False)
teacher_discriminator = teacher.discriminator(
batch_size=teacher_settings["batch_size"])
student_generator = (
model.Registry.models[student_settings["student_network"]]())
hr_size = tf.cast(
tf.convert_to_tensor([1] + student_settings['hr_size']),
tf.float32)
lr_size = tf.cast(hr_size * tf.convert_to_tensor([1, 1 / 4, 1 / 4, 1]),
tf.int32)
logging.debug("Initializing Convolutions")
student_generator.unsigned_call(tf.random.normal(lr_size))
trainer = train.Trainer(teacher_generator,
teacher_discriminator,
summary_writer,
summary_writer_2=teacher_summary_writer,
model_dir=kwargs["modeldir"],
data_dir=kwargs["datadir"],
strategy=strategy)
phase_name = None
if kwargs["type"].lower().startswith("comparative"):
trainer.train_comparative(student_generator,
export_only=stats.get("comparative")
or kwargs["export_only"])
if not kwargs["export_only"]:
stats["comparative"] = True
elif kwargs["type"].lower().startswith("adversarial"):
trainer.train_adversarial(student_generator,
export_only=stats.get("adversarial")
or kwargs["export_only"])
if not kwargs["export_only"]:
stats["adversarial"] = True
# Tracing Graph to put input signature
_ = student_generator.predict(tf.random.normal([1, 180, 320, 3]))
tf.saved_model.save(student_generator,
os.path.join(kwargs["modeldir"], "compressed_esrgan"))
# key = 290
# key = 320
# key = 340
# key = 400
# key = 480
# key = 482
# key = 910
cm = CAMERA_MODEL.CameraModel.load_from_file(
root_dir + dataset + type + sequence + intrinsics_dir + 'intrinsics.txt',
key)
image_loader = IMAGE_LOADER.ImageLoader(root_dir + dataset + type + sequence +
images_dir)
vo = VO_PARSER.VoParserSynth(root_dir, dataset + type + sequence + odom_path)
runtime_settings = SETTINGS.Settings(
10, 0.0001, 0.9, 5, 0.05,
-1.0) # -1.0 for z (or general motion eq.) TODO investigate
ref_count = 8
step = 1
ref_list = []
for count in range(step, ref_count + 1, step):
ref_list.append(key + count)
stereo_pair_builder = STEREO_PAIR_BUILDER.StereoPairBuilder(
cm, image_loader, vo, 0, runtime_settings)
runner = RUNNER.Runner(stereo_pair_builder, runtime_settings, key, ref_list,
root_dir)
# ground_truth = POST.Evaluation.load_ground_truth(depth_file, 424, 512,flip_across_y=True)
# inverted_ground_truth = POST.Evaluation.calc_inverse_ground_truth(ground_truth, runtime_settings,max_thresh=2500,isIpad=True)
def main(**kwargs):
student_settings = settings.Settings("../E2_ESRGAN/config.yaml",
student=True)
teacher_settings = settings.Settings("config/config.yaml", student=False)
root_dir = 'datasets/synth/'
images_dir = root_dir + 'data/img'
odom_path = root_dir + 'info/groundtruth.txt'
mse_dir = root_dir + 'mse/'
key = 1
depth_folder = root_dir + 'data/depth/'
depth_file = IMAGE_LOADER.ImageLoader.parse_id(key) + '.depth'
depth_path = depth_folder + depth_file
cm = CAMERA_MODEL.CameraModel(329.115520046, 329.115520046, 320.0, 240.0)
image_loader = IMAGE_LOADER.ImageLoader(images_dir)
vo = VO_PARSER.VoParserSynth(odom_path)
runtime_settings = SETTINGS.Settings(20, 0.0001, 0.8, 1.0, 0.05, 1.0)
ref_count = 1
ref_list = []
for count in range(1, ref_count + 1):
ref_list.append(key + count)
def load_and_plot_mse():
mse_list = POST.Evaluation.load_data_list(mse_dir + str(key) + '/' +
'mse_list.txt')
name = str(key) + '_mse_line_plot'
VISUALIZE.line_graph(mse_list, root_dir, name)
def load_and_bar_chart_mse():
import sys
import os
from libs import settings
_teacher_directory = settings.Settings(use_student_settings=True)[
"teacher_directory"]
TEACHER_DIR = os.path.abspath(_teacher_directory)
if _teacher_directory not in sys.path:
sys.path.insert(0, TEACHER_DIR)
# key = 1400076116786324 # keyframe
# ref_6 = 1400076139533227
# ref_5 = 1400076139470738
# ref_4 = 1400076139408250
# ref_3 = 1400076139345758
# ref_2 = 1400076139283268
# ref_1 = 1400076139220778
# key = 1400076139158288
camera_loader = CAMERA_MODEL_LOADER.CameraModelLoader(models_dir, cam_type)
cm = camera_loader.load()
image_loader = IMAGE_LOADER.ImageLoader(images_dir)
vo = VO_PARSER.VoParser(odom_path)
runtime_settings = SETTINGS.Settings(cm.focal_length[0], cm.focal_length[1],
cm.principal_point[0],
cm.principal_point[1], 12, 0.01, 0.3, 1.0)
ref_list = [ref_1, ref_2, ref_3]
stereo_pair_builder = STEREO_PAIR_BUILDER.StereoPairBuilder(
cm, image_loader, vo, 25, runtime_settings)
stereo_pairs = stereo_pair_builder.generate_stereo_pairs(key,
ref_list,
invert=True)
iteration = 1
for stereo_pair in stereo_pairs:
#libs.visualize.show_frame(stereo_pair.key_frame.get_image(),root_dir,'keyframe')
it_string = str(iteration)
depth_estimation_re = DEPTH_ESTIMATION.DepthEstimation(
# key = 250
# key = 360
# key = 456
# key = 108
# key = 145
# key = 1152
# cm = CAMERA_MODEL.CameraModel(718.8560000000, 718.8560000000, 607.1928000000, 185.2157000000)
# cm = CAMERA_MODEL.CameraModel(718.8560000000*0.5157131346, 718.8560000000, 320, 185.2157000000)
cm = CAMERA_MODEL.CameraModel(718.8560000000 * 0.5157131346,
718.8560000000 * 1.2765957447, 320,
185.2157000000 * 1.2765957447) #640x480
image_loader = IMAGE_LOADER.ImageLoader(root_dir + images_dir + left_cam_dir)
vo = VO_PARSER.VoParserSynth(root_dir, sequence)
runtime_settings = SETTINGS.Settings(cm.focal_length[0], cm.focal_length[1],
cm.principal_point[0],
cm.principal_point[1], 40, 0.0001, 0.8,
1.0, 0.05)
ref_count = 2
ref_list = []
for count in range(1, ref_count + 1):
ref_list.append(key + count)
stereo_pair_builder = STEREO_PAIR_BUILDER.StereoPairBuilder(
cm, image_loader, vo, 0, runtime_settings)
runner = RUNNER.Runner(stereo_pair_builder, runtime_settings, key, ref_list,
root_dir)
runner.run(VISUALIZE.visualize_enum.SHOW_DEPTH,
normalize=True,
regularize=True,
