def apply_transforms(dataset, confs, batch_size):
"""
Apply transforms to the dataset.
:param confs: A list of configuration like {"name": "scaling", ... }, indicating the specified transform. We use
"--------" to separate the pre-batch-transform and after-batch-transform.
:param batch_size: The batch size to batch the dataset
:return: A new dataset after transform
"""
# Per point cloud instance transform
pre_batch_confs = confs
post_batch_confs = []
try:
batch_seperator_index = [type(x) is str and "---" in x for x in confs].index(True)
pre_batch_confs = confs[:batch_seperator_index]
post_batch_confs = confs[batch_seperator_index + 1:]
except ValueError:
# Cannot find the index
pass
log("Origin dataset={}".format(dataset))
context = {"batch_size": batch_size}
for conf in pre_batch_confs:
transform = object_from_conf(conf, scope="transform", context=context)
dataset = transform(dataset)
log("After pre-batch transform \"{}\" with conf={}, dataset={}".format(conf["name"], conf, dataset))
dataset = dataset.batch(batch_size)
log("Batch transform, dataset={}".format(dataset))
for conf in post_batch_confs:
transform = object_from_conf(conf, scope="transform", context=context)
dataset = transform(dataset)
log("After post-batch transform \"{}\" with conf={}, dataset={}".format(conf["name"], conf, dataset))
return dataset
def learning_rate_from_config(learning_rate_conf):
"""
Get the learning rate scheduler based on configuration
:param learning_rate_conf: The learning rate configuration
:return: A learning rate scheduler
"""
return object_from_conf(learning_rate_conf, scope="learning_rate")
def optimizer_from_config(learning_rate, optimizer_conf):
"""
Get the optimizer from configuration
:param learning_rate: The learning rate, might be a scalar or a learning rate schedule
:param optimizer_conf: The optimizer configuration
:return: An corresponding optimizer
"""
context = {"learning_rate": learning_rate}
return object_from_conf(optimizer_conf, scope="optimizer", context=context)
def layer_from_config(layer_conf, model_conf, data_conf):
"""
Get the corresponding keras layer from configurations
:param layer_conf: The layer configuration
:param model_conf: The global model configuration, sometimes it is used to generate some
special layer like "output-classification" and "output-segmentation" layer
:param data_conf: The dataset configuration, for generating special layers
:return: A keras layer
"""
# context = {"class_count": data_conf["class_count"]}
return object_from_conf(layer_conf, scope="layer", context=None)