Update ForNet.py

ForNet.py

@@ -1126,16 +1126,38 @@ class RecombineDataset(Dataset):
         bg_rat_indices = super()._getitem(0)["bg_rat_idx_file"]
         self.train = "train" in bg_rat_indices.split("/")[-1]
 
-        with open(bg_rat_indices, "r") as f:
-            bg_rat_indices = json.load(f)
-        for in_cls in bg_rat_indices:
-            if in_cls not in self.cls_to_idx:
-                self.cls_to_idx[in_cls] = []
-            for idx, rat in bg_rat_indices[in_cls]:
-                if rat < self.pruning_ratio:
-                    self.cls_to_idx[in_cls].append(idx)
+        bg_rat_idx_file = bg_rat_indices
+        if self.background_combination == "same":
+            try:
+                with open(bg_rat_indices, "r") as f:
+                    bg_rat_indices = json.load(f)
+                for in_cls in bg_rat_indices:
+                    if in_cls not in self.cls_to_idx:
+                        self.cls_to_idx[in_cls] = []
+                    for idx, rat in bg_rat_indices[in_cls]:
+                        if rat < self.pruning_ratio:
+                            self.cls_to_idx[in_cls].append(idx)
+            except (TypeError, KeyError, OSError):
+                logger.warning(
+                    f"Could not load background ratio indices from {bg_rat_indices}. Will do pruning and background selection on the fly. This will take more time in the first few epochs"
+                )
+                self.cls_to_idx = {cls: list(range(len(self))) for cls in IMAGENET2012_CLASSES.keys()}
+
         if self.background_combination == "all":
-            self.cls_to_idx["all"] = list(itertools.chain(*self.cls_to_idx.values()))
+            try:
+                self.cls_to_idx["all"] = []
+                with open(bg_rat_indices, "r") as f:
+                    bg_rat_indices = json.load(f)
+                for in_cls in bg_rat_indices:
+                    for idx, rat in bg_rat_indices[in_cls]:
+                        if rat < self.pruning_ratio:
+                            self.cls_to_idx["all"].append(idx)
+            except (TypeError, KeyError, OSError) as e:
+                logger.warning(f"Error {e} while extracting bg_rat_indices")
+                logger.warning(
+                    f"Could not load background ratio indices from {bg_rat_idx_file}. Will do pruning and background selection on the fly. This will take more time in the first few epochs"
+                )
+                self.cls_to_idx["all"] = list(range(len(self)))
 
     @property
     def total_epochs(self):
@@ -1176,14 +1198,24 @@ class RecombineDataset(Dataset):
         if self.background_combination == "orig":
             bg_item = fg_item
         elif self.background_combination == "same":
-            rand_idx = np.random.randint(len(self.cls_to_idx[in_cls]))
-            rand_idx = self.cls_to_idx[in_cls][rand_idx]
-            bg_item = super()._getitem(rand_idx)
+            while True:
+                rand_idx = np.random.randint(len(self.cls_to_idx[in_cls]))
+                rand_idx = self.cls_to_idx[in_cls][rand_idx]
+                bg_item = super()._getitem(rand_idx)
+                if bg_item["fg/bg_area"] < self.pruning_ratio and bg_item["label"] == fg_item["label"]:
+                    break
+                else:
+                    self.cls_to_idx[in_cls].remove(rand_idx)
         else:
             # all
-            rand_idx = np.random.randint(len(self.cls_to_idx["all"]))
-            rand_idx = self.cls_to_idx["all"][rand_idx]
-            bg_item = super()._getitem(rand_idx)
+            while True:
+                rand_idx = np.random.randint(len(self.cls_to_idx["all"]))
+                rand_idx = self.cls_to_idx["all"][rand_idx]
+                bg_item = super()._getitem(rand_idx)
+                if bg_item["fg/bg_area"] < self.pruning_ratio:
+                    break
+                else:
+                    self.cls_to_idx["all"].remove(rand_idx)
 
         fg_img = fg_item["fg"].convert("RGBA")
         bg_img = bg_item["bg"].convert("RGB")
@@ -1499,6 +1531,7 @@ class ForNet(datasets.GeneratorBasedBuilder):
 
     def _generate_examples(self, patch_files, split, hf_indices, cls_to_idx_loc, fg_bg_ratios):
         logger.info(f"Generating examples from {len(patch_files)} patch files")
+        logger.info(f"will save cls_to_idx to '{cls_to_idx_loc}'")
         logger.info("Opening files")
         class_to_zipfile = {}
         for f in patch_files:
@@ -1590,7 +1623,7 @@ class ForNet(datasets.GeneratorBasedBuilder):
         for proc in zip_procs:
             proc.join()
 
-        # tqdm.write("Finished all processes")
+        tqdm.write("Finished all processes")
         while not ret_queue.empty():
             data = ret_queue.get()
             in_cls = data["path"].split("/")[0]
@@ -1599,7 +1632,7 @@ class ForNet(datasets.GeneratorBasedBuilder):
             cls_to_idx[in_cls].append(foraug_idx)
             yield foraug_idx, data
             foraug_idx += 1
-        tqdm.write("Done")
+        tqdm.write(f"Done generating {split} examples. Saving cls_to_idx file at '{cls_to_idx_loc}'.")
         with open(cls_to_idx_loc, "w") as f:
             json.dump(cls_to_idx, f)
 
@@ -1760,3 +1793,4 @@ def _zip_loader(
                     "fg/bg_area": fg_bg_ratios[patch_name],
                 }
             )
+