def __init__(self, target_name='sigmoid', endpoint_name="sigmoid"):
super().__init__()
self.cross_entropy = nn.BCEWithLogitsLoss(reduce=False)
self.tensorboard_logger = get_tensorboard_logger()
self.logger = get_logger()
self.endpoint_name = endpoint_name
self.target_name = target_name
def __init__(self, endpoint_name, target_name):
super().__init__()
self.endpoint_name = endpoint_name
self.target_name = target_name
self.logger = get_logger()
self.tensorboard_logger = get_tensorboard_logger()
self.l1 = nn.L1Loss(reduction='none')
def __init__(self, alpha, gamma, loss_module, name):
super().__init__(loss_module)
self.alpha = alpha
self.gamma = gamma
self.k = None
self.tensorboard_logger = logger.get_tensorboard_logger()
self.name = name
def __init__(self, target_name, endpoint_name):
super().__init__()
self.cross_entropy = nn.CrossEntropyLoss(reduce=False)
self.tensorboard_logger = get_tensorboard_logger()
self.logger = get_logger()
self.target_name = target_name
self.endpoint_name = endpoint_name
def __init__(self, m, cdist_fn=calc_cdist, endpoint_name='triplet'):
super().__init__()
self.name = "BatchHard(m={})".format(m)
self.m = m
self.cdist_fn = cdist_fn
self.tensorboard_logger = get_tensorboard_logger()
self.logger = get_logger()
self.endpoint_name = endpoint_name
def __init__(self, target_name, endpoint_name, reduction='mean'):
super(CrossEntropyLoss, self).__init__()
self.cross_entropy = nn.CrossEntropyLoss(ignore_index=255, reduction=reduction)
self.tensorboard_logger = get_tensorboard_logger()
#self.logger = get_logger()
self.target_name = target_name
self.endpoint_name = endpoint_name
def __init__(self, loss_module, name, init):
"""
losses (dic): A name-> loss dictionary.
init (float): 0 -> factor is 1
"""
super().__init__(loss_module)
#log(o^2)
tensor = torch.tensor(init, requires_grad=True)
self.log_var = torch.nn.Parameter(tensor, requires_grad=True)
self.name = name
self.logger = logger.get_tensorboard_logger()
def __init__(self, m, cdist_fn=calc_cdist, T=1.0, endpoint_name="triplet"):
"""
Args:
m: margin
T: Softmax temperature
"""
super(BatchSoft, self).__init__()
self.name = "BatchSoft(m={}, T={})".format(m, T)
self.m = m
self.T = T
self.cdist_fn = cdist_fn
self.tensorboard_logger = get_tensorboard_logger()
self.logger = get_logger()
def __init__(self, delta, tr_loss, id_loss):
"""
Args:
tr_loss tuple(name, DynamicFocalLoss): batch hard loss
id_loss tuple(name, DynamicFocalLoss): softmax loss
"""
# TODO
super().__init__({})
# name is name of dataset
self.tr_name, self.tr_loss = tr_loss
self.id_name, self.id_loss = id_loss
self.delta = delta
self.tensorboard_logger = logger.get_tensorboard_logger()
def __init__(self, schedule_fn, optimizer, last_epoch=-1):
"""
Args:
last_epoch: counting from zero
"""
self.schedule_fn = schedule_fn
# Manage underlying pytorch behaviour,
# -1 is special value
if last_epoch == 0:
# from scratch
last_epoch = -1
else:
# restored
last_epoch -= 1
self.last_epoch = last_epoch
self.tensorboard_logger = get_tensorboard_logger()
super().__init__(optimizer, last_epoch)
def __init__(self, target_name, endpoint_name, top_k_percent_pixels=1.0, hard_mining_step=0):
"""
Args:
hard_mining_step: Training step in which the hard mining
kicks off
"""
super().__init__()
self.cross_entropy = nn.CrossEntropyLoss(ignore_index=255, reduction='none')
self.tensorboard_logger = get_tensorboard_logger()
#self.logger = get_logger()
self.target_name = target_name
self.endpoint_name = endpoint_name
if top_k_percent_pixels == 1.0:
# just default cross entropy loss
self.forward = super().forward
self.top_k_percent_pixels = top_k_percent_pixels
self.hard_mining_step = hard_mining_step
# TODO global step
self.step = 0
def __init__(self, attributes):
super().__init__()
self.attributes = attributes
self.cross_entropy = nn.CrossEntropyLoss(reduce=False)
self.M = len(attributes)
self.tensorboard_logger = get_tensorboard_logger()
def __init__(self, endpoint_name, target_name):
super().__init__()
self.endpoint_name = endpoint_name
self.target_name = target_name
self.logger = get_logger()
self.tensorboard_logger = get_tensorboard_logger()
def __init__(self, losses):
super().__init__()
self.losses = torch.nn.ModuleDict(losses)
self.logger = logger.get_tensorboard_logger()
def __init__(self, tasks, model, scale_only_backbone=False):
super().__init__()
self.model = model
self.tasks = tasks
self.tb = logger.get_tensorboard_logger()
self.scale_only_backbone = scale_only_backbone
