def make_loss(args, gids):
"""
Construct loss function(s).
"""
gid = None if gids is None else gids[0]
if args.loss_type == 'softmax':
criterion = CrossEntropyLoss()
elif args.loss_type == 'contrastive':
criterion = ContrastiveLoss(margin=args.margin)
elif args.loss_type == 'triplet':
criterion = TripletLoss(margin=args.margin)
elif args.loss_type == 'softmax-triplet':
criterion = {
'softmax': CrossEntropyLoss(),
'triplet': TripletLoss(margin=args.margin)
}
elif args.loss_type == 'npair':
criterion = NpairLoss(reg_lambda=0.002, gid=gid)
elif args.loss_type == 'lifted':
criterion = LiftedLoss(margin=args.margin, gid=gid)
elif args.loss_type == 'dmml':
criterion = DMMLLoss(num_support=args.num_support,
distance_mode=args.distance_mode,
margin=args.margin,
gid=gid)
else:
raise NotImplementedError
return criterion
def __init__(self, args):
super(gradcam_Loss, self).__init__()
print('[INFO] Making gradcam_Loss...')
self.nGPU = args.nGPU
self.args = args
self.loss = []
self.loss_module = nn.ModuleList()
for loss in args.gradcam_loss.split('+'):
weight, loss_type = loss.split('*')
if loss_type == 'CrossEntropy':
loss_function = nn.CrossEntropyLoss()
elif loss_type == 'Triplet':
loss_function = TripletLoss(args.margin)
self.loss.append({
'type': loss_type,
'weight': float(weight),
'function': loss_function
})
if len(self.loss) > 1:
self.loss.append({'type': 'Total', 'weight': 0, 'function': None})
self.device = torch.device('cuda')
self.loss_module.to(self.device)
if args.nGPU > 1:
self.loss_module = nn.DataParallel(
self.loss_module, range(args.nGPU)
)
def __init__(self, config):
self.config = config
self.device = self._get_device()
self.train_loader = self._load_lvis_results()
if self.config['loss']['type'] == 'nce':
from loss.nt_xent import NTXentLoss
self.loss_crit = NTXentLoss(self.device, config['batch_size'],
**config['loss'])
if self.config['loss']['include_hierarchical']:
self.hierarchical_loss_crit = HierarchicalLoss(
margin=config['loss']['margin'])
if self.config['hyperbolic']:
self.triplet_loss_crit = HTripletLoss(
margin=config['loss']['margin'])
else:
self.triplet_loss_crit = TripletLoss(
margin=config['loss']['margin'])
def __init__(self, args, ckpt):
super(Loss, self).__init__()
print('[INFO] Making loss...')
self.nGPU = args.nGPU
self.args = args
self.loss = []
self.loss_module = nn.ModuleList()
device = torch.device('cpu' if args.cpu else 'cuda')
for loss in args.loss.split('+'):
weight, loss_type = loss.split('*')
if loss_type == 'CrossEntropy':
loss_function = nn.CrossEntropyLoss()
elif loss_type == 'Triplet':
loss_function = TripletLoss(args.margin)
elif loss_type == 'TripletSemihard':
loss_function = TripletSemihardLoss(device, args.margin)
elif loss_type == 'Center':
loss_function = CenterLoss(device,
args.num_classes,
feat_dim=256)
self.loss.append({
'type': loss_type,
'weight': float(weight),
'function': loss_function
})
if len(self.loss) > 1:
self.loss.append({'type': 'Total', 'weight': 0, 'function': None})
for l in self.loss:
if l['function'] is not None:
print('{:.3f} * {}'.format(l['weight'], l['type']))
self.loss_module.append(l['function'])
self.log = torch.Tensor()
self.loss_module.to(device)
if args.load != '': self.load(ckpt.dir, cpu=args.cpu)
if not args.cpu and args.nGPU > 1:
self.loss_module = nn.DataParallel(self.loss_module,
range(args.nGPU))
def __init__(self, args, ckpt):
super(Loss, self).__init__()
print('[INFO] Making loss...')
self.nGPU = args.nGPU
self.args = args
self.loss = []
self.loss_module = nn.ModuleList()
self.circular_mixed_loss_queue = None
mixed_subloss_options = {
'TripletSemihard': {
'function':
TripletSemihardLoss(
torch.device('cpu' if args.cpu else 'cuda'), args.margin),
'subtype':
'TripletSemihard'
},
'Triplet': {
'function': TripletLoss(args.margin),
'subtype': 'Triplet'
}
}
subtype = ''
for loss in args.loss.split('+'):
weight, loss_type = loss.split('*')
if loss_type == 'CrossEntropy':
loss_function = nn.CrossEntropyLoss()
elif loss_type == 'Triplet':
loss_function = TripletLoss(args.margin)
# CSCE 625: Aligned loss evaluation for aligned features
elif loss_type == 'AlignedTriplet':
tri_loss = TripletLoss2(margin=0.3)
loss_function = AlignedTripletLoss(tri_loss)
# ------------ BELOW CODE FOR CSCE 625 ---------------
# Allow a mixed loss function for the training...
# switch out loss functions after a fixed number of epochs
# set in args
elif loss_type.startswith('Mixed'):
print("Will cycle loss functions {} every {} epochs."\
.format(repr(loss_type.split('-')[1:]), self.args.switch_loss_every))
self.circular_mixed_loss_queue = cycle(
mixed_subloss_options[l] for l in loss_type.split('-')[1:])
# also append other loss functions to be used in the mix
l = next(self.circular_mixed_loss_queue)
loss_function = l['function']
subtype = l['subtype']
loss_type = 'Mixed'
# ----------------------------------------------------
self.loss.append({
'type': loss_type,
'weight': float(weight),
'function': loss_function,
'subtype': subtype
})
subtype = ''
# CSCE 625: Mutual Learning
if args.mutual_learning:
self.ml_pm_weight = 1.0
self.ml_global_weight = 0.0
self.ml_local_weight = 1.0
self.loss.append({
'type': 'ProbabilityML',
'weight': 1.0,
'function': None
})
self.loss.append({
'type': 'GlobalML',
'weight': 0.0, # update after implementation
'function': None
})
self.loss.append({
'type': 'LocalML',
'weight': 1.0,
'function': None
})
if len(self.loss) > 1:
self.loss.append({'type': 'Total', 'weight': 0, 'function': None})
for l in self.loss:
if l['function'] is not None:
print('{:.3f} * {}'.format(l['weight'], l['type']))
self.loss_module.append(l['function'])
self.device = torch.device('cpu' if args.cpu else 'cuda')
self.loss_module.to(self.device)
if args.load != '': self.load(ckpt.dir, cpu=args.cpu)
if not args.cpu and args.nGPU > 1:
self.loss_module = nn.DataParallel(self.loss_module,
range(args.nGPU))
self.log = [torch.Tensor()]
if args.mutual_learning:
self.log.append(torch.Tensor())