support registering IterableDataset and build it from config.

Summary:
Fix facebookresearch#4105

Pull Request resolved: facebookresearch#4180

Reviewed By: zhanghang1989

Differential Revision: D36045130

fbshipit-source-id: f2aa3bfdebf476737b6deec8eac93ef3043964b8

INSTALL.md

@@ -187,7 +187,8 @@ C++ compilation errors from NVCC / NVRTC, or "Unsupported gpu architecture"
 <br/>
 A few possibilities:
 
-1. Local CUDA/NVCC version has to match the CUDA version of your PyTorch. Both can be found in `python collect_env.py`.
+1. Local CUDA/NVCC version has to match the CUDA version of your PyTorch. Both can be found in `python collect_env.py`
+   (download from [here](./detectron2/utils/collect_env.py)).
    When they are inconsistent, you need to either install a different build of PyTorch (or build by yourself)
    to match your local CUDA installation, or install a different version of CUDA to match PyTorch.
 

detectron2/data/build.py

@@ -239,6 +239,15 @@ def get_detection_dataset_dicts(
         names = [names]
     assert len(names), names
     dataset_dicts = [DatasetCatalog.get(dataset_name) for dataset_name in names]
+
+    if isinstance(dataset_dicts[0], torchdata.Dataset):
+        if len(dataset_dicts) > 1:
+            # ConcatDataset does not work for iterable style dataset.
+            # We could support concat for iterable as well, but it's often
+            # not a good idea to concat iterables anyway.
+            return torchdata.ConcatDataset(dataset_dicts)
+        return dataset_dicts[0]
+
     for dataset_name, dicts in zip(names, dataset_dicts):
         assert len(dicts), "Dataset '{}' is empty!".format(dataset_name)
 
@@ -250,9 +259,6 @@ def get_detection_dataset_dicts(
             for dataset_i_dicts, proposal_file in zip(dataset_dicts, proposal_files)
         ]
 
-    if isinstance(dataset_dicts[0], torchdata.Dataset):
-        return torchdata.ConcatDataset(dataset_dicts)
-
     dataset_dicts = list(itertools.chain.from_iterable(dataset_dicts))
 
     has_instances = "annotations" in dataset_dicts[0]
@@ -351,18 +357,24 @@ def _train_loader_from_config(cfg, mapper=None, *, dataset=None, sampler=None):
     if sampler is None:
         sampler_name = cfg.DATALOADER.SAMPLER_TRAIN
         logger = logging.getLogger(__name__)
-        logger.info("Using training sampler {}".format(sampler_name))
-        if sampler_name == "TrainingSampler":
-            sampler = TrainingSampler(len(dataset))
-        elif sampler_name == "RepeatFactorTrainingSampler":
-            repeat_factors = RepeatFactorTrainingSampler.repeat_factors_from_category_frequency(
-                dataset, cfg.DATALOADER.REPEAT_THRESHOLD
-            )
-            sampler = RepeatFactorTrainingSampler(repeat_factors)
-        elif sampler_name == "RandomSubsetTrainingSampler":
-            sampler = RandomSubsetTrainingSampler(len(dataset), cfg.DATALOADER.RANDOM_SUBSET_RATIO)
+        if isinstance(dataset, torchdata.IterableDataset):
+            logger.info("Not using any sampler since the dataset is IterableDataset.")
+            sampler = None
         else:
-            raise ValueError("Unknown training sampler: {}".format(sampler_name))
+            logger.info("Using training sampler {}".format(sampler_name))
+            if sampler_name == "TrainingSampler":
+                sampler = TrainingSampler(len(dataset))
+            elif sampler_name == "RepeatFactorTrainingSampler":
+                repeat_factors = RepeatFactorTrainingSampler.repeat_factors_from_category_frequency(
+                    dataset, cfg.DATALOADER.REPEAT_THRESHOLD
+                )
+                sampler = RepeatFactorTrainingSampler(repeat_factors)
+            elif sampler_name == "RandomSubsetTrainingSampler":
+                sampler = RandomSubsetTrainingSampler(
+                    len(dataset), cfg.DATALOADER.RANDOM_SUBSET_RATIO
+                )
+            else:
+                raise ValueError("Unknown training sampler: {}".format(sampler_name))
 
     return {
         "dataset": dataset,
@@ -461,7 +473,9 @@ def _test_loader_from_config(cfg, dataset_name, mapper=None):
         "dataset": dataset,
         "mapper": mapper,
         "num_workers": cfg.DATALOADER.NUM_WORKERS,
-        "sampler": InferenceSampler(len(dataset)),
+        "sampler": InferenceSampler(len(dataset))
+        if not isinstance(dataset, torchdata.IterableDataset)
+        else None,
     }
 
 

tests/data/test_dataset.py

@@ -9,8 +9,9 @@
 from iopath.common.file_io import LazyPath
 
 from detectron2 import model_zoo
-from detectron2.config import instantiate
+from detectron2.config import get_cfg, instantiate
 from detectron2.data import (
+    DatasetCatalog,
     DatasetFromList,
     MapDataset,
     ToIterableDataset,
@@ -112,6 +113,22 @@ def test_build_iterable_dataloader_train(self):
         dl = build_detection_train_loader(dataset=ds, **kwargs)
         next(iter(dl))
 
+    def test_build_iterable_dataloader_from_cfg(self):
+        cfg = get_cfg()
+
+        class MyData(torch.utils.data.IterableDataset):
+            def __iter__(self):
+                while True:
+                    yield 1
+
+        cfg.DATASETS.TRAIN = ["iter_data"]
+        DatasetCatalog.register("iter_data", lambda: MyData())
+        dl = build_detection_train_loader(cfg, mapper=lambda x: x, aspect_ratio_grouping=False)
+        next(iter(dl))
+
+        dl = build_detection_test_loader(cfg, "iter_data", mapper=lambda x: x)
+        next(iter(dl))
+
     def _check_is_range(self, data_loader, N):
         # check that data_loader produces range(N)
         data = list(iter(data_loader))