def build_final_activation(self, model_config=None):
model_config = self.get(
'model') if model_config is None else model_config
model_class = list(model_config.keys())[0]
final_layer = None
final_activation_config = model_config[model_class].pop(
'final_activation', True)
if not final_activation_config:
return
elif final_activation_config is True:
final_activation_config = {}
for layer_name, kwargs in final_activation_config.items():
layer_class = locate(layer_name, [
'embeddingutils.models.submodules',
'inferno.extensions.layers.convolutional'
])
final_layer = layer_class(**kwargs)
model_config[model_class]['final_activation'] = \
Conv3D(in_channels=model_config[model_class]['out_channels'],
out_channels=model_config[model_class]['out_channels'],
kernel_size=3)
if final_layer is not None:
model_config[model_class]['final_activation'] = nn.Sequential(
model_config[model_class]['final_activation'], final_layer)
def build_final_activation(self, model_config=None):
model_config = self.get(
'model') if model_config is None else model_config
model_class = list(model_config.keys())[0]
final_activation = model_config[model_class].pop(
'final_activation', None)
if final_activation is None:
return
final_activation = locate(final_activation, ['torch.nn'])
model_config[model_class]['final_activation'] = \
final_activation()
def inferno_build_metric(self):
metric_config = self.get('trainer/metric')
frequency = metric_config.pop('evaluate_every', (25, 'iterations'))
self.trainer.evaluate_metric_every(frequency)
if metric_config:
assert len(metric_config) == 1
for class_name, kwargs in metric_config.items():
cls = locate(class_name)
kwargs['offsets'] = self.get('global/offsets')
print(f'Building metric of class "{cls.__name__}"')
metric = cls(**kwargs)
self.trainer.build_metric(metric)
self.set('trainer/metric/evaluate_every', frequency)
def parse_and_wrap_losses(self, config, transforms, losses, weights,
loss_names):
default_weight = config.pop('weight', 1)
for class_name, kwargs in config.items():
loss_names.append(kwargs.pop('name', class_name))
weights.append(kwargs.pop('weight', default_weight))
print(f'Adding {loss_names[-1]} with weight {weights[-1]}')
loss_class = locate(class_name, [
'embeddingutils.loss', 'SegTags.loss',
'inferno.extensions.criteria.set_similarity_measures',
'torch.nn'
])
if issubclass(loss_class, WeightedLoss):
kwargs['trainer'] = self.trainer
losses.append(
LossWrapper(criterion=loss_class(**kwargs),
transforms=transforms))
def inferno_build_metric(self):
metric_config = self.get('trainer/metric')
frequency = metric_config.pop('evaluate_every', (25, 'iterations'))
self.trainer.evaluate_metric_every(frequency)
if metric_config:
assert len(metric_config) == 1
for class_name, kwargs in metric_config.items():
cls = locate(class_name)
if issubclass(cls, SFRPMetric):
kwargs['trainer'] = self.trainer
print(f'Building metric of class "{cls.__name__}"')
metric = cls(**kwargs)
if hasattr(self, 'metric_pre'):
pre = self.metric_pre()
self.trainer.build_metric(lambda prediction, target:
metric(*pre(prediction, target)))
else:
self.trainer.build_metric(metric)
self.set('trainer/metric/evaluate_every', frequency)