add DATALOADER.REPEAT_SQRT

Summary:
Pull Request resolved: #5245

X-link: fairinternal/detectron2#602

For sampler **RepeatFactorTrainingSampler**, current per-category weight is computed as **1/sqrt(frequency)**.

This works fine on LVIS but is not sufficient in highly imbalanced data we have for person segmentation.

Thus we add an argument **DATALOADER.REPEAT_SQRT**. If false, we compute per-category weight as **1/frequency**.

This change is entirely back-compatible.

Reviewed By: wat3rBro

Differential Revision:
D55355021

Privacy Context Container: L1165023

fbshipit-source-id: 6bca2eecc3b9a7b4693a288c5779627254cd5ec5

detectron2/config/defaults.py

@@ -121,6 +121,8 @@
 _C.DATALOADER.SAMPLER_TRAIN = "TrainingSampler"
 # Repeat threshold for RepeatFactorTrainingSampler
 _C.DATALOADER.REPEAT_THRESHOLD = 0.0
+# if True, take square root when computing repeating factor
+_C.DATALOADER.REPEAT_SQRT = True
 # Tf True, when working on datasets that have instance annotations, the
 # training dataloader will filter out images without associated annotations
 _C.DATALOADER.FILTER_EMPTY_ANNOTATIONS = True

detectron2/data/build.py

@@ -430,7 +430,7 @@ def _build_weighted_sampler(cfg, enable_category_balance=False):
         """
         category_repeat_factors = [
             RepeatFactorTrainingSampler.repeat_factors_from_category_frequency(
-                dataset_dict, cfg.DATALOADER.REPEAT_THRESHOLD
+                dataset_dict, cfg.DATALOADER.REPEAT_THRESHOLD, sqrt=cfg.DATALOADER.REPEAT_SQRT
             )
             for dataset_dict in dataset_name_to_dicts.values()
         ]
@@ -482,7 +482,7 @@ def _train_loader_from_config(cfg, mapper=None, *, dataset=None, sampler=None):
                 sampler = TrainingSampler(len(dataset))
             elif sampler_name == "RepeatFactorTrainingSampler":
                 repeat_factors = RepeatFactorTrainingSampler.repeat_factors_from_category_frequency(
-                    dataset, cfg.DATALOADER.REPEAT_THRESHOLD
+                    dataset, cfg.DATALOADER.REPEAT_THRESHOLD, sqrt=cfg.DATALOADER.REPEAT_SQRT
                 )
                 sampler = RepeatFactorTrainingSampler(repeat_factors)
             elif sampler_name == "RandomSubsetTrainingSampler":

detectron2/data/samplers/distributed_sampler.py

@@ -155,7 +155,7 @@ def __init__(self, repeat_factors, *, shuffle=True, seed=None):
         self._frac_part = repeat_factors - self._int_part
 
     @staticmethod
-    def repeat_factors_from_category_frequency(dataset_dicts, repeat_thresh):
+    def repeat_factors_from_category_frequency(dataset_dicts, repeat_thresh, sqrt=True):
         """
         Compute (fractional) per-image repeat factors based on category frequency.
         The repeat factor for an image is a function of the frequency of the rarest
@@ -169,6 +169,7 @@ def repeat_factors_from_category_frequency(dataset_dicts, repeat_thresh):
             repeat_thresh (float): frequency threshold below which data is repeated.
                 If the frequency is half of `repeat_thresh`, the image will be
                 repeated twice.
+            sqrt (bool): if True, apply :func:`math.sqrt` to the repeat factor.
 
         Returns:
             torch.Tensor:
@@ -187,7 +188,14 @@ def repeat_factors_from_category_frequency(dataset_dicts, repeat_thresh):
         # 2. For each category c, compute the category-level repeat factor:
         #    r(c) = max(1, sqrt(t / f(c)))
         category_rep = {
-            cat_id: max(1.0, math.sqrt(repeat_thresh / cat_freq))
+            cat_id: max(
+                1.0,
+                (
+                    math.sqrt(repeat_thresh / cat_freq)
+                    if sqrt
+                    else (repeat_thresh / cat_freq)
+                ),
+            )
             for cat_id, cat_freq in category_freq.items()
         }
         for cat_id in sorted(category_rep.keys()):