## Description This is the part 2 of Gluon Data API extension and fixes, which mainly focus on speed up the current data loading pipeline using gluon dataset and dataloader.
## Motivation The current data loading pipeline is the major bottleneck for many training tasks. We can summarize the entire flow as: ```bash | Dataset.__getitem__ -> | Transform.__call__()/forward() -> | Batchify -> | (optional communicate through shared_mem) -> | split_and_load(ctxs) -> | <training on GPUs> -> ``` where there are performance concerns: - performance of python dataset/transform functions aren't satisfying - it's not easy to embrace multithreading to speed up dataloading due to global interpreter lock - python multiprocessing is unfortunately slow and error prune, not to mention the shared memory implementations on different OS are quite difference and very annoying(e.g., it's very likely to run out of shared memory if not properly taken care of) - currently memory planing for batchify is non-exist, causing frequent alloc/dealloc for large chunk of memory if the batch size is big - batchify then split and load can be optimized to partial_batchify ## Proposal To alleviate the existing troubles I propose to use a hybrid solution, that is to - provide C++ Datasets that can cover the most usecases ```python from gluon.data.dataset import TupleDataset, ImageFolderDataset, ArrayDataset # as long as TupleDataset, ImageSequenceDataset, ArrayDataset are supported by backend dataset = TupleDataset([ImageSequenceDataset(img_paths), ArrayDataset(image_labels)]) # dataset is an image classification dataset while fully supported in C++ # with TupleDataset we can combine as many data as possible # a C++ backed Dataset can have a magic __handle__ method to return the c++ handle for reference class TupleDataset: def __init__(self, datasets): if all([callable(getattr(dataset, '__handle__')) for dataset in datasets]): # all supported by backend self._tuple_dataset = check_call(_LIB.MXTupleDatasetCreate([getattr(dataset, '__handle__') for dataset in datasets])) else: self._tuple_dataset = None def __handle__(self): return self._tuple_dataset ``` - provide common C++ batchify functions that are split and context aware. Batchify with memory planner is TBD. - provide a C++ `MultithreadingDataLoader` which inherit the same arguments as `gluon.data.DataLoader` but use mxnet internal multithreading rather than python multiprocessing. - fallback to python multiprocessing whenever - the dataset is not fully supported by backend(e.g., there are custom python datasets) - Transform is not fully hybridizable - Batchify is not fully supported by backend User will continue to use the existing `gluon.data.DataLoader`, and the conversion will be applied automatically ```python loader = gluon.data.DataLoader(hybrid_dataset.transform(hybrid_transform), batch_size=32, batchify_fn=hybrid_batchify) def DataLoader: def __init__(self, dataset, ...): if isinstance(dataset, _LazyTransformDataset) and is_hybrid(dataset._transform) and is_hybrid(dataset) and is_hybrid(batchify_fn): self._mt_dataloader = check_call(_LIB.MXMultiThreadDataLoaderCreate(...)) def __iter__(self): if self._mt_dataloader: return self._mt_dataloader else: # fallback to single thread normal dataloader or multiprocessing dataloader ``` With this change, mxnet 2.0 will get smooth transition to mixed data loaders. Please comment with specific examples where this proposal fail to accommodate. -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: https://github.com/apache/incubator-mxnet/issues/17269