update logger and trainer (clip regularization if excessive)

data/coco-stuff/make_annotation.py

@@ -9,17 +9,17 @@
 
 for split in ["train2017", "val2017"]:
     annotations = f"{orig_path}/annotations/{split}"
-    nov_ann = f"{dest_path}/annotations_my/{split}"
+    nov_ann = f"{dest_path}/annotations_obj/{split}"
 
     # clear folder if exists
     if osp.exists(nov_ann):
         print("Removing existing")
         os.rmdir(nov_ann)
     os.makedirs(nov_ann)
 
-    # remap labels in the novel interval (+1 for Stuff, +1 and stuff on 0 for objects)
+    # remap labels in the novel interval (+1 for Stuff, +1 excluding stuff for objects)
     mapping = np.zeros((256,), dtype=np.int8)
-    for i, cl in enumerate(range(91)):
+    for i, cl in enumerate(range(182)):
         mapping[cl] = i + 1
     mapping[255] = 255
     target_transform = lambda x: Image.fromarray(mapping[x])

data/coco/make_annotation.py

@@ -4,22 +4,22 @@
 import os.path as osp
 from tqdm import tqdm
 
-orig_path = ''  # insert here the source path where original COCO annotations are
+orig_path = '.'  # insert here the source path where original COCO annotations are
 dest_path = '.'  # the destination folder, which should be this one'
 
 for split in ["train2017", "val2017"]:
     annotations = f"{orig_path}/annotations/{split}"
-    nov_ann = f"{dest_path}/annotations_obj/{split}"
+    nov_ann = f"{dest_path}/annotations_my/{split}"
 
     # clear folder if exists
     if osp.exists(nov_ann):
         print("Removing existing")
         os.rmdir(nov_ann)
     os.makedirs(nov_ann)
 
-    # remap labels in the novel interval (+1 for Stuff, +1 excluding stuff for objects)
+    # remap labels in the novel interval (+1 for Stuff, +1 and stuff on 0 for objects)
     mapping = np.zeros((256,), dtype=np.int8)
-    for i, cl in enumerate(range(182)):
+    for i, cl in enumerate(range(91)):
         mapping[cl] = i + 1
     mapping[255] = 255
     target_transform = lambda x: Image.fromarray(mapping[x])

methods/trainer.py

@@ -8,6 +8,7 @@
 from modules.classifier import IncrementalClassifier, CosineClassifier, SPNetClassifier
 from .utils import get_scheduler, get_batch, MeanReduction
 
+CLIP = 10
 
 class Trainer:
     def __init__(self, task, device, logger, opts):
@@ -211,10 +212,13 @@ def train(self, cur_epoch, train_loader, metrics=None, print_int=10, n_iter=1):
 
                 loss = self.reduction(criterion(outputs, labels), labels)
 
-                loss_tot = loss + rloss
-                assert not torch.isnan(loss_tot), "Error, loss is NaN!"
-                loss_tot.backward()
+                if rloss <= CLIP:
+                    loss_tot = loss + rloss
+                else:
+                    print(f"Warning, rloss is {rloss}! Term ignored Skipped")
+                    loss_tot = loss
 
+                loss_tot.backward()
                 optim.step()
                 scheduler.step()
 

utils/logger.py

@@ -94,7 +94,7 @@ def error(self, msg):
 
     def log_results(self, task, name, results, novel=False):
         if self.rank == 0:
-            file_name = f"{task.task}.csv"
+            file_name = f"{task.task}-n{task.nshot}.csv"
             file_name = file_name if not novel else f"{file_name}_novel.csv"
             dir_path = f"{self.logdir_results}/{task.dataset}"
             path = f"{dir_path}/{file_name}"