def create(
self,
arch,
num_output_channels,
num_input_channels,
loss,
lr,
optimizer,
lrsch,
momentum=0.9,
weight_decay=5e-4,
pretrained=False,
size_input=388,
num_classes=8,
backbone='preactresnet',
num_filters=32,
breal='real',
alpha=2,
beta=2,
):
"""
Create
-arch (string): architecture
-loss (string):
-lr (float): learning rate
-optimizer (string) :
-lrsch (string): scheduler learning rate
-pretrained (bool)
"""
cfg_opt = {'momentum': momentum, 'weight_decay': weight_decay}
#cfg_scheduler={ 'step_size':100, 'gamma':0.1 }
cfg_scheduler = {'mode': 'min', 'patience': 10}
cfg_model = {'num_filters': num_filters}
self.num_classes = num_classes
super(ClassNeuralNet, self).create(
arch,
num_output_channels,
num_input_channels,
loss,
lr,
optimizer,
lrsch,
pretrained,
cfg_opt=cfg_opt,
cfg_scheduler=cfg_scheduler,
cfg_model=cfg_model,
)
self.size_input = size_input
self.backbone = backbone
self.num_filters = num_filters
self.topk = nloss.TopkAccuracy()
self.logger_train = Logger('Train', ['loss', 'loss_bce'], ['topk'],
self.plotter)
self.logger_val = Logger('Val ', ['loss', 'loss_bce'], ['topk'],
self.plotter)
self.breal = breal
def create(
self,
arch,
num_output_channels,
num_input_channels,
loss,
lr,
optimizer,
lrsch,
momentum=0.9,
weight_decay=5e-4,
pretrained=False,
topk=(1, ),
size_input=128,
):
"""
Create
Args:
arch (string): architecture
num_output_channels,
num_input_channels,
loss (string):
lr (float): learning rate
momentum,
optimizer (string) :
lrsch (string): scheduler learning rate
pretrained (bool)
"""
cfg_opt = {'momentum': 0.9, 'weight_decay': 5e-4}
cfg_scheduler = {'step_size': 100, 'gamma': 0.1}
super(NeuralNetClassifier, self).create(
arch,
num_output_channels,
num_input_channels,
loss,
lr,
optimizer,
lrsch,
pretrained,
cfg_opt=cfg_opt,
cfg_scheduler=cfg_scheduler,
)
self.size_input = size_input
self.accuracy = nloss.TopkAccuracy(topk)
self.cnf = nloss.ConfusionMeter(self.num_output_channels,
normalized=True)
self.visheatmap = gph.HeatMapVisdom(env_name=self.nameproject)
# Set the graphic visualization
self.metrics_name = ['top{}'.format(k) for k in topk]
self.logger_train = Logger('Trn', ['loss'], self.metrics_name,
self.plotter)
self.logger_val = Logger('Val', ['loss'], self.metrics_name,
self.plotter)
def create(
self,
arch,
num_output_channels,
num_input_channels,
loss,
lr,
optimizer,
lrsch,
momentum=0.9,
weight_decay=5e-4,
pretrained=False,
th=0.0,
size_input=32,
):
"""
Create
Args:
arch (string): architecture
num_output_channels,
num_input_channels,
loss (string):
lr (float): learning rate
momentum,
optimizer (string) :
lrsch (string): scheduler learning rate
pretrained (bool)
"""
cfg_opt = {'momentum': 0.9, 'weight_decay': 5e-4}
cfg_scheduler = {'step_size': 50, 'gamma': 0.1}
super(NeuralNetClassifier, self).create(
arch,
num_output_channels,
num_input_channels,
loss,
lr,
optimizer,
lrsch,
pretrained,
cfg_opt=cfg_opt,
cfg_scheduler=cfg_scheduler,
)
self.size_input = size_input
self.accuracy = nloss.MultAccuracyV1(th)
self.f_score = nloss.F_score(threshold=th, beta=2)
#self.cnf = nloss.ConfusionMeter( self.num_output_channels, normalized=True )
#self.visheatmap = gph.HeatMapVisdom( env_name=self.nameproject )
# Set the graphic visualization
self.logger_train = Logger('Trn', ['loss'], ['acc', 'f1'],
self.plotter)
self.logger_val = Logger('Val', ['loss'], ['acc', 'f1'], self.plotter)
def create(self,
arch,
num_output_channels,
num_input_channels,
loss,
lr,
optimizer,
lrsch,
momentum=0.9,
weight_decay=5e-4,
pretrained=False,
size_input=388,
num_classes=8,
):
"""
Create
Args:
-arch (string): architecture
-num_output_channels,
-num_input_channels,
-loss (string):
-lr (float): learning rate
-optimizer (string) :
-lrsch (string): scheduler learning rate
-pretrained (bool)
-
"""
super(AttentionGMMNeuralNet, self).create(
arch,
num_output_channels,
num_input_channels,
loss,
lr,
optimizer,
lrsch,
momentum,
weight_decay,
pretrained,
size_input,
num_classes,
)
self.logger_train = Logger( 'Train', ['loss', 'loss_gmm', 'loss_bce', 'loss_att' ], [ 'topk', 'gmm'], self.plotter )
self.logger_val = Logger( 'Val ', ['loss', 'loss_gmm', 'loss_bce', 'loss_att' ], [ 'topk', 'gmm'], self.plotter )
def create(
self,
arch,
num_output_channels,
num_input_channels,
loss,
lr,
momentum,
optimizer,
lrsch,
pretrained=False,
topk=(1, ),
):
"""
Create
Args:
@arch (string): architecture
@num_output_channels,
@num_input_channels,
@loss (string):
@lr (float): learning rate
@momentum,
@optimizer (string) :
@lrsch (string): scheduler learning rate
@pretrained (bool)
"""
super(NeuralNetClassifier,
self).create(arch, num_output_channels, num_input_channels, loss,
lr, momentum, optimizer, lrsch, pretrained)
self.accuracy = nloss.Accuracy(topk)
self.cnf = nloss.ConfusionMeter(self.num_output_channels,
normalized=True)
self.visheatmap = gph.HeatMapVisdom(env_name=self.nameproject)
# Set the graphic visualization
self.metrics_name = ['top{}'.format(k) for k in topk]
self.logger_train = Logger('Trn', ['loss'], self.metrics_name,
self.plotter)
self.logger_val = Logger('Val', ['loss'], self.metrics_name,
self.plotter)
def create(
self,
arch,
num_output_channels,
num_input_channels,
loss,
lr,
momentum,
optimizer,
lrsch,
pretrained=False,
size_input=388,
):
"""
Create
-arch (string): architecture
-loss (string):
-lr (float): learning rate
-optimizer (string) :
-lrsch (string): scheduler learning rate
-pretrained (bool)
"""
super(SegmentationNeuralNet,
self).create(arch, num_output_channels, num_input_channels, loss,
lr, momentum, optimizer, lrsch, pretrained)
self.size_input = size_input
self.accuracy = nloss.Accuracy()
self.dice = nloss.Dice()
# Set the graphic visualization
self.logger_train = Logger('Train', ['loss'], ['accs', 'dices'],
self.plotter)
self.logger_val = Logger('Val ', ['loss'], ['accs', 'dices'],
self.plotter)
self.visheatmap = gph.HeatMapVisdom(env_name=self.nameproject,
heatsize=(100, 100))
self.visimshow = gph.ImageVisdom(env_name=self.nameproject,
imsize=(100, 100))
def create(self,
arch,
num_output_channels,
num_input_channels,
loss,
lr,
optimizer,
lrsch,
momentum=0.9,
weight_decay=5e-4,
pretrained=False,
size_input=388,
cascade_type='none'):
"""
Create
Args:
-arch (string): architecture
-num_output_channels,
-num_input_channels,
-loss (string):
-lr (float): learning rate
-optimizer (string) :
-lrsch (string): scheduler learning rate
-pretrained (bool)
"""
cfg_opt = {'momentum': momentum, 'weight_decay': weight_decay}
cfg_scheduler = {'step_size': 100, 'gamma': 0.1}
super(SegmentationNeuralNet, self).create(arch,
num_output_channels,
num_input_channels,
loss,
lr,
optimizer,
lrsch,
pretrained,
cfg_opt=cfg_opt,
cfg_scheduler=cfg_scheduler)
self.size_input = size_input
self.num_output_channels = num_output_channels
self.cascade_type = cascade_type
self.segs_per_forward = 7
if self.cascade_type == 'none':
self.step = self.default_step
elif self.cascade_type == 'ransac':
self.step = self.ransac_step
elif self.cascade_type == 'ransac2':
self.step = self.ransac_step2
elif self.cascade_type == 'simple':
self.step = self.cascate_step
else:
raise "Cascada not found"
self.accuracy = nloss.Accuracy()
if num_output_channels == 2:
dice_dim = (1, )
if num_output_channels == 4:
dice_dim = (1, 2, 3)
self.dice = nloss.Dice(dice_dim)
# Set the graphic visualization
self.logger_train = Logger('Train', ['loss'], ['accs', 'dices'],
self.plotter)
self.logger_val = Logger('Val ', ['loss'], ['accs', 'dices', 'PQ'],
self.plotter)
self.visheatmap = gph.HeatMapVisdom(env_name=self.nameproject,
heatsize=(256, 256))
self.visimshow = gph.ImageVisdom(env_name=self.nameproject,
imsize=(256, 256))
if self.half_precision:
self.scaler = torch.cuda.amp.GradScaler()