train : added sample_weights option (to be tested)

deepfinder/training.py

@@ -166,6 +166,7 @@ class Train(core.DeepFinder):
         self.Lrnd = 13  # random shifts applied when sampling data- and target-patches (in voxels)
 
         self.class_weight = None
+        self.sample_weights = None  # np array same lenght as objl_train
 
         self.check_attributes()
 
@@ -261,8 +262,16 @@ class Train(core.DeepFinder):
                 if self.flag_direct_read:
                     batch_data, batch_target = self.generate_batch_direct_read(path_data, path_target, self.batch_size, objlist_train)
                 else:
-                    batch_data, batch_target = self.generate_batch_from_array(data_list, target_list, self.batch_size, objlist_train)
-                loss_train = self.net.train_on_batch(batch_data, batch_target, class_weight=self.class_weight)
+                    batch_data, batch_target, idx_list = self.generate_batch_from_array(data_list, target_list, self.batch_size, objlist_train)
+
+                if self.sample_weights is not None:
+                    sample_weight = self.sample_weights[idx_list]
+                else:
+                    sample_weight = None
+
+                loss_train = self.net.train_on_batch(batch_data, batch_target,
+                                                     class_weight=self.class_weight,
+                                                     sample_weight=sample_weight)
 
                 self.display('epoch %d/%d - it %d/%d - loss: %0.3f - acc: %0.3f' % (e + 1, self.epochs, it + 1, self.steps_per_epoch, loss_train[0], loss_train[1]))
                 list_loss_train.append(loss_train[0])
@@ -280,7 +289,7 @@ class Train(core.DeepFinder):
                 if self.flag_direct_read:
                     batch_data_valid, batch_target_valid = self.generate_batch_direct_read(path_data, path_target, self.batch_size, objlist_valid)
                 else:
-                    batch_data_valid, batch_target_valid = self.generate_batch_from_array(data_list, target_list, self.batch_size, objlist_valid)
+                    batch_data_valid, batch_target_valid, idx_list = self.generate_batch_from_array(data_list, target_list, self.batch_size, objlist_valid)
                 loss_val = self.net.evaluate(batch_data_valid, batch_target_valid, verbose=0) # TODO replace by loss() to reduce computation
                 batch_pred = self.net.predict(batch_data_valid)
                 #loss_val = K.eval(losses.tversky_loss(K.constant(batch_target_valid), K.constant(batch_pred)))
@@ -417,9 +426,11 @@ class Train(core.DeepFinder):
         else:  # choose from whole objlist
             pool = range(0, len(objlist))
 
+        idx_list = []
         for i in range(batch_size):
             # choose random sample in training set:
             index = np.random.choice(pool)
+            idx_list.append(index)
 
             tomoID = int(objlist[index]['tomo_idx'])
 
@@ -448,4 +459,5 @@ class Train(core.DeepFinder):
                 batch_data[i] = np.rot90(batch_data[i], k=2, axes=(0, 2))
                 batch_target[i] = np.rot90(batch_target[i], k=2, axes=(0, 2))
 
-        return batch_data, batch_target
+        return batch_data, batch_target, idx_list
\ No newline at end of file
+