def build_model(self):
# backbone encoder
backbone, skips = self.get_backbone()
# decoder construction
prediction = self.create_decoder(backbone, skips)
network = keras.Model(inputs=backbone.input,
outputs=prediction,
name="FPN_NET")
network.summary()
# get the optimizer
optimizer = self.build_optimizer()
metrics = self.build_metrics_NYU()
if self.config.model.loss == "focal_loss":
loss = CategoricalFocalLoss(gamma=self.config.model.gamma,
alpha=self.config.model.alpha)
elif self.config.model.loss == "lovasz":
loss = MultiClassLovaszSoftmaxLoss()
else:
loss = 'categorical_crossentropy'
network.compile(loss=loss, optimizer=optimizer, metrics=metrics)
return network
def build_model(self):
# backbone encoder
backbone, skips = self.get_backbone()
# joint decoder (resolution 1/4)
decoder = self.decoder(backbone, skips)
# segmentation head
segmentation_mask = self.segmentation_head(decoder)
# depth estimation head
depth_estimation = self.depth_head(decoder)
# create model
network = keras.Model(inputs=backbone.input,
outputs=[segmentation_mask, depth_estimation],
name="JOINT_BI_FPN")
network.summary()
# get the optimizer
optimizer = self.build_optimizer()
# define segmentation loss
if self.config.model.seg_loss == "focal_loss":
seg_loss = CategoricalFocalLoss(gamma=self.config.model.gamma,
alpha=self.config.model.alpha)
elif self.config.model.seg_loss == "lovasz":
seg_loss = MultiClassLovaszSoftmaxLoss()
else:
seg_loss = 'categorical_crossentropy'
# define depth loss
if self.config.model.depth_loss == "Huber":
depth_loss = tf.keras.losses.Huber()
else:
depth_loss = tf.keras.losses.MeanSquaredError()
# composed loss
losses = {"seg_out": seg_loss, "dep_out": depth_loss}
loss_weights = {
"seg_out": self.config.model.loss_weights.seg_loss,
"dep_out": self.config.model.loss_weights.depth_loss
}
# define metrics
metrics = {
"seg_out": self._generate_metrics(),
"dep_out": [tf.keras.metrics.RootMeanSquaredError()]
}
network.compile(loss=losses,
loss_weights=loss_weights,
optimizer=optimizer,
metrics=metrics)
return network
def segmentation_loss(self):
# define segmentation loss
if self.config.model.seg_loss == "focal_loss":
seg_loss = CategoricalFocalLoss(gamma=self.config.model.gamma,
alpha=self.config.model.alpha)
elif self.config.model.seg_loss == "lovasz":
seg_loss = MultiClassLovaszSoftmaxLoss()
elif self.config.model.seg_loss == "focal_tversky":
seg_loss = focal_tversky_loss
elif self.config.model.seg_loss == "tversky":
seg_loss = class_tversky
else:
seg_loss = 'categorical_crossentropy'
return seg_loss