Train: added self.loss argument. This makes it more easy to load custom losses.

deepfinder/training.py

@@ -158,6 +158,7 @@ class Train(core.DeepFinder):
         self.steps_per_epoch = 100
         self.steps_per_valid = 10  # number of samples for validation
         self.optimizer = Adam(lr=0.0001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
+        self.loss = losses.tversky_loss
 
         self.flag_direct_read = 1
         self.flag_batch_bootstrap = 0
@@ -231,7 +232,7 @@ class Train(core.DeepFinder):
 
 
         # Build network (not in constructor, else not possible to init model with weights from previous train round):
-        self.net.compile(optimizer=self.optimizer, loss=losses.tversky_loss, metrics=['accuracy'])
+        self.net.compile(optimizer=self.optimizer, loss=self.loss, metrics=['accuracy'])
 
         # Load whole dataset:
         if self.flag_direct_read == False: