[Fix] Replace value.numpy()[0] with float(value) (PaddlePaddle#2686)

* To support 0 dim Tensor, replace value.numpy()[0] with float(value) when the shape of value is [1]

Matting/ppmatting/core/train.py

@@ -227,7 +227,7 @@ def train(model,
             model.clear_gradients()
 
             for key, value in loss_dict.items():
-                avg_loss[key] += value.numpy()[0]
+                avg_loss[key] += float(value)
             batch_cost_averager.record(
                 time.time() - batch_start, num_samples=batch_size)
 

Matting/ppmatting/core/val_ml.py

@@ -44,10 +44,10 @@ def reverse_transform(alpha, trans_info):
     """recover pred to origin shape"""
     for item in trans_info[::-1]:
         if item[0][0] == 'resize':
-            h, w = item[1][0].numpy()[0], item[1][1].numpy()[0]
+            h, w = int(item[1][0]), int(item[1][1])
             alpha = cv2.resize(alpha, dsize=(w, h))
         elif item[0][0] == 'padding':
-            h, w = item[1][0].numpy()[0], item[1][1].numpy()[0]
+            h, w = int(item[1][0]), int(item[1][1])
             alpha = alpha[0:h, 0:w]
         else:
             raise Exception("Unexpected info '{}' in im_info".format(item[0]))

contrib/AutoNUE/scripts/train.py

@@ -167,7 +167,7 @@ def train(model,
                           paddle.optimizer.lr.LRScheduler):
                 optimizer._learning_rate.step()
             model.clear_gradients()
-            avg_loss += loss.numpy()[0]
+            avg_loss += float(loss)
             if not avg_loss_list:
                 avg_loss_list = [l.numpy() for l in loss_list]
             else:

contrib/CityscapesSOTA/scripts/train.py

@@ -170,7 +170,7 @@ def train(model,
                           paddle.optimizer.lr.LRScheduler):
                 optimizer._learning_rate.step()
             model.clear_gradients()
-            avg_loss += loss.numpy()[0]
+            avg_loss += float(loss)
             if not avg_loss_list:
                 avg_loss_list = [l.numpy() for l in loss_list]
             else:

contrib/DomainAdaptation/script/train.py

@@ -177,8 +177,8 @@ def train(self,
                                                      labels_src)
 
                 loss_src_seg = loss_src_seg_main + loss_src_seg_aux
-                loss_dict["source_main"] = loss_src_seg_main.numpy()[0]
-                loss_dict["source_aux"] = loss_src_seg_aux.numpy()[0]
+                loss_dict["source_main"] = float(loss_src_seg_main)
+                loss_dict["source_aux"] = float(loss_src_seg_aux)
                 loss = loss_src_seg
                 del loss_src_seg, loss_src_seg_aux, loss_src_seg_main
 
@@ -224,8 +224,8 @@ def train(self,
 
                     loss += loss_edge
 
-                    loss_dict['target_edge'] = loss_tgt_edge.numpy()[0]
-                    loss_dict['source_edge'] = loss_src_edge.numpy()[0]
+                    loss_dict['target_edge'] = float(loss_tgt_edge)
+                    loss_dict['source_edge'] = float(loss_src_edge)
 
                     del loss_edge, loss_tgt_edge, loss_src_edge
 
@@ -260,8 +260,8 @@ def train(self,
 
                 loss += loss_tgt_aug
 
-                loss_dict['target_aug_main'] = loss_tgt_aug_main.numpy()[0]
-                loss_dict['target_aug_aux'] = loss_tgt_aug_aux.numpy()[0]
+                loss_dict['target_aug_main'] = float(loss_tgt_aug_main)
+                loss_dict['target_aug_aux'] = float(loss_tgt_aug_aux)
                 del images_tgt_aug, labels_tgt_aug_aux, images_tgt, \
                     loss_tgt_aug, loss_tgt_aug_aux, loss_tgt_aug_main
 
@@ -287,8 +287,8 @@ def train(self,
                     loss_edge_rec = loss_tgt_edge_rec + loss_src_edge_rec
                     loss += loss_edge_rec
 
-                    loss_dict['src_edge_rec'] = loss_src_edge_rec.numpy()[0]
-                    loss_dict['tgt_edge_rec'] = loss_tgt_edge_rec.numpy()[0]
+                    loss_dict['src_edge_rec'] = float(loss_src_edge_rec)
+                    loss_dict['tgt_edge_rec'] = float(loss_tgt_edge_rec)
 
                     del loss_tgt_edge_rec, loss_src_edge_rec
 
@@ -358,12 +358,12 @@ def train(self,
 
                         loss += loss_feat_align
 
-                        loss_dict['loss_pix_align_src'] = \
-                                loss_pix_align_src.numpy()[0]
-                        loss_dict['loss_pix_align_tgt'] = \
-                                loss_pix_align_tgt.numpy()[0]
-                        loss_dict['loss_intra_relate'] = \
-                                loss_intra_relate.numpy()[0]
+                        loss_dict['loss_pix_align_src'] = float(
+                            loss_pix_align_src)
+                        loss_dict['loss_pix_align_tgt'] = float(
+                            loss_pix_align_tgt)
+                        loss_dict['loss_intra_relate'] = float(
+                            loss_intra_relate)
 
                         del loss_pix_align_tgt, loss_pix_align_src, loss_intra_relate,
 

contrib/LaneSeg/core/train.py

@@ -210,7 +210,7 @@ def train(model,
             train_profiler.add_profiler_step(profiler_options)
 
             model.clear_gradients()
-            avg_loss += loss.numpy()[0]
+            avg_loss += float(loss)
             if not avg_loss_list:
                 avg_loss_list = [l.numpy() for l in loss_list]
             else:

contrib/MedicalSeg/medicalseg/core/train.py

@@ -161,7 +161,7 @@ def train(model,
             model.clear_gradients()
             # TODO use a function to record, print lossetc
 
-            avg_loss += loss.numpy()[0]
+            avg_loss += float(loss)
             mdice += np.mean(per_channel_dice) * 100
 
             if channel_dice_array.size == 0:

contrib/MedicalSeg/nnunet/core/train.py

@@ -141,7 +141,7 @@ def train(model,
             train_profiler.add_profiler_step(profiler_options)
 
             model.clear_gradients()
-            avg_loss += loss.numpy()[0]
+            avg_loss += float(loss)
             mdice += np.mean(per_channel_dice) * 100
 
             if channel_dice_array.size == 0:

contrib/PanopticDeepLab/core/train.py

@@ -189,7 +189,7 @@ def train(model,
                           paddle.optimizer.lr.LRScheduler):
                 optimizer._learning_rate.step()
             model.clear_gradients()
-            avg_loss += loss.numpy()[0]
+            avg_loss += float(loss)
             if not avg_loss_list:
                 avg_loss_list = [l.numpy() for l in loss_list]
             else:

deploy/slim/distill/distill_utils.py

@@ -244,10 +244,10 @@ def distill_train(distill_model,
                 lr_sche.step()
 
             distill_model.clear_gradients()
-            avg_loss += loss.numpy()[0]
-            avg_out_loss += out_loss.numpy()[0]
-            avg_out_distill_loss += out_distill_loss.numpy()[0]
-            avg_feature_distill_loss += feature_distill_loss.numpy()[0]
+            avg_loss += float(loss)
+            avg_out_loss += float(out_loss)
+            avg_out_distill_loss += float(out_distill_loss)
+            avg_feature_distill_loss += float(feature_distill_loss)
             if not avg_out_loss_list:
                 avg_out_loss_list = [l.numpy() for l in out_loss_list]
             else:

paddleseg/core/train.py

@@ -230,7 +230,7 @@ def train(model,
             train_profiler.add_profiler_step(profiler_options)
 
             model.clear_gradients()
-            avg_loss += loss.numpy()[0]
+            avg_loss += float(loss)
             if not avg_loss_list:
                 avg_loss_list = [l.numpy() for l in loss_list]
             else:

paddleseg/models/losses/ohem_cross_entropy_loss.py

@@ -77,7 +77,7 @@ def forward(self, logit, label):
             if self.min_kept > 0:
                 index = prob.argsort()
                 threshold_index = index[min(len(index), self.min_kept) - 1]
-                threshold_index = int(threshold_index.numpy()[0])
+                threshold_index = int(threshold_index)
                 if prob[threshold_index] > self.thresh:
                     threshold = prob[threshold_index]
                 kept_mask = (prob < threshold).astype('int64')

paddleseg/models/losses/ohem_edge_attention_loss.py

@@ -93,7 +93,7 @@ def forward(self, logits, label):
             if self.min_kept > 0:
                 index = prob.argsort()
                 threshold_index = index[min(len(index), self.min_kept) - 1]
-                threshold_index = int(threshold_index.numpy()[0])
+                threshold_index = int(threshold_index)
                 if prob[threshold_index] > self.thresh:
                     threshold = prob[threshold_index]
                 kept_mask = (prob < threshold).astype('int64')