Add README for DataLoader2 (#526)

Summary:
Add brief description of `DataLoader2` and related components.

Pull Request resolved: #526

Reviewed By: NivekT

Differential Revision: D37315170

Pulled By: ejguan

fbshipit-source-id: 9886fb60c48870815c4e8242eb46e69031abba9a

torchdata/dataloader2/README.md

@@ -0,0 +1,85 @@
+# DataLoader2 (Prototype)
+
+`DataLoader2` is introduced to execute `DataPipe` graph, with support for
+
+- dynamic sharding for multi-process and distributed data loading
+- multiple backend `ReadingServices` (WIP for `DistributedReadingService`)
+- `DataPipe` graph in-place modification like shuffle control, memory pinning, etc.
+- snapshot the state of data-preprocessing pipeline (WIP)
+
+## Why DataLoader2
+
+`DataPipe` is introduced to not only decompose the data pre-processing operations, but also decouple the overloaded
+data-manipulation features from `DataLoader`. So, a light-weight `DataLoader2` is required. Besides, a certain features
+can only be achieved with `DataLoader2` like snapshotting and switching backend services to perform high-performant
+operations.
+
+These options are configured by the constructor arguments of `DataLoader2`, which has the signature:
+
+```py
+DataLoader2(datapipe, datapipe_adapter_fn=None, reading_service=None)
+```
+
+The sections below describe in details the effects and usages of these options.
+
+## DataPipe
+
+`DataLoader2` supports both `MapDataPipe` and `IterDataPipe`. To understand the basic structure of `DataPipe`, please
+see [What are DataPipes?](https://github.com/pytorch/data#what-are-datapipes).
+
+Note: `DataLoader2` doesn't support `Dataset` or `IterableDataset`. Please wrap each of them with the corresponding
+`DataPipe`.
+
+- [`SequenceWrapper`](https://pytorch.org/data/main/generated/torchdata.datapipes.map.SequenceWrapper.html#sequencewrapper):
+  `Dataset`
+- [`IterableWrapper`](https://pytorch.org/data/main/generated/torchdata.datapipes.iter.IterableWrapper.html#iterablewrapper):
+  `IterableDataset`
+
+## Data-Processing Options (`datapipe_adapter_fn`)
+
+`Adapters` are used to configure, modify and extend the graph of `DataPipes` in `DataLoader2`. It allows in-place
+modification on the pre-assembled graph of `DataPipes` provided by PyTorch domains. For example, `Shuffle(False)` can be
+provided to `DataLoader2`, which would disable any of `shuffle` operations in the graph of `DataPipes`.
+
+And, we will provide more `Adapters` to cover data-processing options (WIP):
+
+- `PinMemory`: Attach a `DataPipe` at the end of the data-processing graph that coverts output data to `torch.Tensor` in
+  pinned memory.
+- `FullSync`: Attach a `DataPipe` to make sure the data-processing graph synchronized between distributed processes to
+  prevent hanging.
+- `ShardingPolicy`: Modify sharding policy if `sharding_filter` is presented in the graph of `DataPipes`.
+- `PrefetchPolicy`, `InvalidateCache`, etc.
+
+If you have feature requests about the `Adapters` you'd like to be provided, please open a GitHub issue. For specific
+needs, `DataLoader2` also accepts any custom `Adapter` as long as it inherits from the `Adapter` class.
+
+## ReadingService
+
+`ReadingService` specifies the execution backend for the data-processing graph. There are two types of `ReadingServices`
+in TorchData:
+
+- `MultiprocessingReadingService`
+- `DistributedReadingService`
+
+These two `ReadingServices` would take the graph of `DataPipes` and modify it to achieve a few features like dynamic
+sharding, sharing seeds and snapshoting for multi-/distributed processes.
+
+This also allows easier transition of data-preprocessing pipeline from research to production. After the graph of
+`DataPipes` is created and validated with the `ReadingServices`, a different `ReadingService` that configures and
+connects to the production service/infra such as `AIStore` can be provided to `DataLoader2` as a drop-in replacement.
+This `ReadingService` could potentially search the graph, and find `DataPipe` operations that can be delegated to the
+production service/infra, then modify the graph correspondingly to achieve higher-performant execution.
+
+A sequence of graph utilities are provided to help users to define their own `ReadingService` and modify the graph:
+
+- `torchdata.dataloader2.graph.traverse(datapipe)`: Get graph of `DataPipes`
+- `torchdata.dataloader2.graph.find_dps(graph, datapipe_type)`: Find `DataPipe` based on the type from the `DataPipe`
+  graph
+- `torchdata.dataloader2.graph.replace_dp(graph, old_datapipe, new_datapipe)`: Replace a `DataPipe` by another
+  `DataPipe` in the `DataPipe` graph
+- `torchdata.dataloader2.graph.remove_dp(graph, datapipe)`: Remove a `DataPipe` from the `DataPipe` graph
+
+## Prototype Usage and Feedback
+
+`DataLoader2` is stable in terms of API, but functionally not complete yet. We welcome early adopters and feedback, as
+well as potential contributors.