TNauen commited on
Commit
a763881
·
verified ·
1 Parent(s): b798fe6

Update ForNet.py

Browse files
Files changed (1) hide show
  1. ForNet.py +51 -17
ForNet.py CHANGED
@@ -1126,16 +1126,38 @@ class RecombineDataset(Dataset):
1126
  bg_rat_indices = super()._getitem(0)["bg_rat_idx_file"]
1127
  self.train = "train" in bg_rat_indices.split("/")[-1]
1128
 
1129
- with open(bg_rat_indices, "r") as f:
1130
- bg_rat_indices = json.load(f)
1131
- for in_cls in bg_rat_indices:
1132
- if in_cls not in self.cls_to_idx:
1133
- self.cls_to_idx[in_cls] = []
1134
- for idx, rat in bg_rat_indices[in_cls]:
1135
- if rat < self.pruning_ratio:
1136
- self.cls_to_idx[in_cls].append(idx)
 
 
 
 
 
 
 
 
 
1137
  if self.background_combination == "all":
1138
- self.cls_to_idx["all"] = list(itertools.chain(*self.cls_to_idx.values()))
 
 
 
 
 
 
 
 
 
 
 
 
 
1139
 
1140
  @property
1141
  def total_epochs(self):
@@ -1176,14 +1198,24 @@ class RecombineDataset(Dataset):
1176
  if self.background_combination == "orig":
1177
  bg_item = fg_item
1178
  elif self.background_combination == "same":
1179
- rand_idx = np.random.randint(len(self.cls_to_idx[in_cls]))
1180
- rand_idx = self.cls_to_idx[in_cls][rand_idx]
1181
- bg_item = super()._getitem(rand_idx)
 
 
 
 
 
1182
  else:
1183
  # all
1184
- rand_idx = np.random.randint(len(self.cls_to_idx["all"]))
1185
- rand_idx = self.cls_to_idx["all"][rand_idx]
1186
- bg_item = super()._getitem(rand_idx)
 
 
 
 
 
1187
 
1188
  fg_img = fg_item["fg"].convert("RGBA")
1189
  bg_img = bg_item["bg"].convert("RGB")
@@ -1499,6 +1531,7 @@ class ForNet(datasets.GeneratorBasedBuilder):
1499
 
1500
  def _generate_examples(self, patch_files, split, hf_indices, cls_to_idx_loc, fg_bg_ratios):
1501
  logger.info(f"Generating examples from {len(patch_files)} patch files")
 
1502
  logger.info("Opening files")
1503
  class_to_zipfile = {}
1504
  for f in patch_files:
@@ -1590,7 +1623,7 @@ class ForNet(datasets.GeneratorBasedBuilder):
1590
  for proc in zip_procs:
1591
  proc.join()
1592
 
1593
- # tqdm.write("Finished all processes")
1594
  while not ret_queue.empty():
1595
  data = ret_queue.get()
1596
  in_cls = data["path"].split("/")[0]
@@ -1599,7 +1632,7 @@ class ForNet(datasets.GeneratorBasedBuilder):
1599
  cls_to_idx[in_cls].append(foraug_idx)
1600
  yield foraug_idx, data
1601
  foraug_idx += 1
1602
- tqdm.write("Done")
1603
  with open(cls_to_idx_loc, "w") as f:
1604
  json.dump(cls_to_idx, f)
1605
 
@@ -1760,3 +1793,4 @@ def _zip_loader(
1760
  "fg/bg_area": fg_bg_ratios[patch_name],
1761
  }
1762
  )
 
 
1126
  bg_rat_indices = super()._getitem(0)["bg_rat_idx_file"]
1127
  self.train = "train" in bg_rat_indices.split("/")[-1]
1128
 
1129
+ bg_rat_idx_file = bg_rat_indices
1130
+ if self.background_combination == "same":
1131
+ try:
1132
+ with open(bg_rat_indices, "r") as f:
1133
+ bg_rat_indices = json.load(f)
1134
+ for in_cls in bg_rat_indices:
1135
+ if in_cls not in self.cls_to_idx:
1136
+ self.cls_to_idx[in_cls] = []
1137
+ for idx, rat in bg_rat_indices[in_cls]:
1138
+ if rat < self.pruning_ratio:
1139
+ self.cls_to_idx[in_cls].append(idx)
1140
+ except (TypeError, KeyError, OSError):
1141
+ logger.warning(
1142
+ 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"
1143
+ )
1144
+ self.cls_to_idx = {cls: list(range(len(self))) for cls in IMAGENET2012_CLASSES.keys()}
1145
+
1146
  if self.background_combination == "all":
1147
+ try:
1148
+ self.cls_to_idx["all"] = []
1149
+ with open(bg_rat_indices, "r") as f:
1150
+ bg_rat_indices = json.load(f)
1151
+ for in_cls in bg_rat_indices:
1152
+ for idx, rat in bg_rat_indices[in_cls]:
1153
+ if rat < self.pruning_ratio:
1154
+ self.cls_to_idx["all"].append(idx)
1155
+ except (TypeError, KeyError, OSError) as e:
1156
+ logger.warning(f"Error {e} while extracting bg_rat_indices")
1157
+ logger.warning(
1158
+ 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"
1159
+ )
1160
+ self.cls_to_idx["all"] = list(range(len(self)))
1161
 
1162
  @property
1163
  def total_epochs(self):
 
1198
  if self.background_combination == "orig":
1199
  bg_item = fg_item
1200
  elif self.background_combination == "same":
1201
+ while True:
1202
+ rand_idx = np.random.randint(len(self.cls_to_idx[in_cls]))
1203
+ rand_idx = self.cls_to_idx[in_cls][rand_idx]
1204
+ bg_item = super()._getitem(rand_idx)
1205
+ if bg_item["fg/bg_area"] < self.pruning_ratio and bg_item["label"] == fg_item["label"]:
1206
+ break
1207
+ else:
1208
+ self.cls_to_idx[in_cls].remove(rand_idx)
1209
  else:
1210
  # all
1211
+ while True:
1212
+ rand_idx = np.random.randint(len(self.cls_to_idx["all"]))
1213
+ rand_idx = self.cls_to_idx["all"][rand_idx]
1214
+ bg_item = super()._getitem(rand_idx)
1215
+ if bg_item["fg/bg_area"] < self.pruning_ratio:
1216
+ break
1217
+ else:
1218
+ self.cls_to_idx["all"].remove(rand_idx)
1219
 
1220
  fg_img = fg_item["fg"].convert("RGBA")
1221
  bg_img = bg_item["bg"].convert("RGB")
 
1531
 
1532
  def _generate_examples(self, patch_files, split, hf_indices, cls_to_idx_loc, fg_bg_ratios):
1533
  logger.info(f"Generating examples from {len(patch_files)} patch files")
1534
+ logger.info(f"will save cls_to_idx to '{cls_to_idx_loc}'")
1535
  logger.info("Opening files")
1536
  class_to_zipfile = {}
1537
  for f in patch_files:
 
1623
  for proc in zip_procs:
1624
  proc.join()
1625
 
1626
+ tqdm.write("Finished all processes")
1627
  while not ret_queue.empty():
1628
  data = ret_queue.get()
1629
  in_cls = data["path"].split("/")[0]
 
1632
  cls_to_idx[in_cls].append(foraug_idx)
1633
  yield foraug_idx, data
1634
  foraug_idx += 1
1635
+ tqdm.write(f"Done generating {split} examples. Saving cls_to_idx file at '{cls_to_idx_loc}'.")
1636
  with open(cls_to_idx_loc, "w") as f:
1637
  json.dump(cls_to_idx, f)
1638
 
 
1793
  "fg/bg_area": fg_bg_ratios[patch_name],
1794
  }
1795
  )
1796
+