[SGD] Dataset API (#7839)

doc/source/index.rst

@@ -181,6 +181,7 @@ Getting Involved
    raysgd/raysgd.rst
    raysgd/raysgd_pytorch.rst
    raysgd/raysgd_tensorflow.rst
+   raysgd/raysgd_dataset.rst
    raysgd/raysgd_ref.rst
 
 .. toctree::

doc/source/raysgd/raysgd_dataset.rst

@@ -0,0 +1,48 @@
+Distributed Dataset
+===================
+
+The RaySGD ``Dataset`` provides a simple abstraction for training with
+distributed data.
+
+.. tip:: Get in touch with us if you're using or considering using `RaySGD <https://forms.gle/26EMwdahdgm7Lscy9>`_!
+
+Setting up a dataset
+--------------------
+
+A dataset can be constructed via any python iterable, or a ``ParallelIterator``. Optionally, a batch size, download function, concurrency, and a transformation can also be specified.
+
+When constructing a dataset, a download function can be specified. For example, if a dataset is initialized with a set of paths, a download function can be specified which converts those paths to ``(input, label)`` tuples. The download function can be executed in parallel via ``max_concurrency``. This may be useful if the backing datastore has rate limits, there is high overhead associated with a download, or downloading is computationally expensive. Downloaded data is stored as objects in the plasma store. 
+
+An additional, final transformation can be specified via ``Dataset::transform``. This function is guaranteed to take place on the same worker that training will take place on. It is good practice to do operations which produce large outputs, such as converting images to tensors as transformations.
+
+Finally, the batch size can be specified. The batch size is the number of data points used per training step per worker.
+
+.. note:: Batch size should be specified via the dataset's constructor, __not__ the ``config["batch_size"]`` passed into the Trainer constructor. In general, datasets are configured via their own constructor, not the Trainer config, wherever possible. 
+
+Using a dataset
+---------------
+
+To use a dataset, pass it in as an argument to ``trainer.train()``. A dataset passed in to ``trainer.train`` will take precedence over the trainer's data creator during that training run.
+
+.. code-block:: python
+
+    trainer.train(dataset=dataset, num_steps=10) # Trains using a dataset
+    trainer.train() # Trains with the original data creator
+    trainer.train(dataset=dataset2, num_steps=20) # Trains using a different dataset
+
+Sharding and Sampling
+---------------------
+
+.. note:: These details may change in the future. 
+
+Datasets use ParallelIterator actors for sharding. In order to handle datasets which do not shard evenly, and streaming datasets (which may not have a defined size), shards are represented as repeated sequences of data. As a result, num_steps should always be specified when training and some data may be oversampled if the data cannot be evenly sharded.
+
+If the dataset is of a known length (and can be evenly sharded), training for an epoch is eqivalent to setting ``num_steps = len(data) / (num_workers * batch_size)``. 
+
+Complete dataset example
+------------------------
+
+Below is an example of training a network with a single hidden layer to learn the identity function.
+
+.. literalinclude:: ../../../python/ray/util/sgd/data/examples/mlp_identity.py
+   :language: python

doc/source/raysgd/raysgd_ref.rst

@@ -33,3 +33,12 @@ TFTrainer
     :members:
 
     .. automethod:: __init__
+
+Dataset
+-------
+
+.. autoclass:: ray.util.sgd.data.Dataset
+    :members:
+
+    .. automethod:: __init__
+

python/ray/util/sgd/data/__init__.py

@@ -0,0 +1,5 @@
+from ray.util.sgd.data.dataset import Dataset
+import logging
+logger = logging.getLogger(__name__)
+
+__all__ = ["Dataset"]

python/ray/util/sgd/data/dataset.py

@@ -0,0 +1,92 @@
+from ray.util.iter import ParallelIterator, from_iterators
+
+
+class Dataset():
+    """A simple Dataset abstraction for RaySGD.
+
+    This dataset is designed to work with RaySGD trainers (currently just
+    Torch) to provide support for streaming large external datasets, and built
+    in sharding.
+
+    .. code-block:: python
+
+        def to_mat(x):
+            return torch.tensor([[x]]).float()
+
+
+        data = [i * 0.001 for i in range(1000)]
+        p_iter = iter.from_items(data, num_shards=1, repeat=True)
+        dataset = Dataset(
+            p_iter,
+            batch_size=32,
+            max_concurrency=1,
+            download_func=lambda x: (to_mat(x), to_mat(x)))
+
+        trainer = TorchTrainer(
+            model_creator=model_creator,
+            data_creator=None,
+            optimizer_creator=optimizer_creator,
+            loss_creator=torch.nn.MSELoss,
+            num_workers=5,
+        )
+
+        for i in range(10):
+            # Train for another epoch using the dataset
+            trainer.train(dataset=dataset, num_steps=200)
+
+        model = trainer.get_model()
+        print("f(0.5)=", float(model(to_mat(0.5))[0][0]))
+
+    Args:
+        data (iterable[U] or ParallelIterator[U]): Any existing python
+            iterable (or iterator), or an existing parallel iterator
+            to use.
+        batch_size (int): The batch size for training/inference (default 32).
+        download_func (U -> (S, Y)): A function which returns two values, the
+            input and the label (default is the identity function).
+        max_concurrency (int): The maximum number of concurrent calls to the
+            download function. See ParallelIterator::for_each for details.
+        transform (S -> X): A final transformation to be applied to the _input
+            only_. This is guaranteed to run on the same worker that training
+            will occur on.
+    """
+
+    def __init__(self,
+                 data,
+                 batch_size=32,
+                 download_func=None,
+                 max_concurrency=0,
+                 transform=None):
+        par_iter = None
+        if isinstance(data, ParallelIterator):
+            par_iter = data.repartition(1)
+        else:
+            par_iter = from_iterators([data], repeat=True)
+        if download_func:
+            par_iter = par_iter.for_each(
+                download_func, max_concurrency=max_concurrency)
+        self.iter = par_iter.batch(batch_size)
+
+        self.batch_size = batch_size
+        self.max_concurrency = max_concurrency
+        self.transform = transform
+
+    def set_num_shards(self, num_shards):
+        """
+        Reshards the iterator if necessary.
+        """
+        if num_shards != self.iter.num_shards():
+            print("Setting num shards", num_shards)
+            self.iter = self.iter.repartition(num_shards)
+
+    def get_shard(self, i):
+        """
+        Returns a single, iterable shard.
+        """
+        assert i < self.iter.num_shards(), \
+            "Trying to get shard {} but there are only {} shards." + \
+            "Are you sure you called set_num_shards already?".format(
+                i, self.iter.num_shards()
+            )
+
+        return self.iter.get_shard(i)

python/ray/util/sgd/data/examples/.gitignore

@@ -0,0 +1 @@
+ray-project/*

python/ray/util/sgd/data/examples/mlp_identity.py

@@ -0,0 +1,69 @@
+import ray
+from ray.util.sgd.torch.torch_trainer import TorchTrainer
+from ray.util.sgd.data.dataset import Dataset
+
+import torch
+from torch import nn
+import torch.nn.functional as F
+
+
+class Net(nn.Module):
+    def __init__(self):
+        super(Net, self).__init__()
+        self.fc1 = nn.Linear(1, 128)
+        self.fc2 = nn.Linear(128, 1)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = F.relu(x)
+        x = self.fc2(x)
+        return x
+
+
+def model_creator(config):
+    return Net()
+
+
+def optimizer_creator(model, config):
+    return torch.optim.SGD(model.parameters(), lr=config.get("lr", 1e-4))
+
+
+def to_mat(x):
+    return torch.tensor([[x]]).float()
+
+
+def dataset_creator():
+    num_points = 32 * 100 * 2
+    data = [i * (1 / num_points) for i in range(num_points)]
+    dataset = Dataset(
+        data,
+        batch_size=32,
+        max_concurrency=2,
+        download_func=lambda x: (to_mat(x), to_mat(x)))
+    return dataset
+
+
+def main():
+    dataset = dataset_creator()
+    trainer = TorchTrainer(
+        model_creator=model_creator,
+        data_creator=None,
+        optimizer_creator=optimizer_creator,
+        loss_creator=torch.nn.MSELoss,
+        num_workers=2,
+    )
+
+    for i in range(10):
+        # Train a full epoch using the data_creator
+        # trainer.train()
+
+        # Train for another epoch using the dataset
+        trainer.train(dataset=dataset, num_steps=100)
+
+        model = trainer.get_model()
+        print("f(0.5)=", float(model(to_mat(0.5))[0][0]))
+
+
+if __name__ == "__main__":
+    ray.init()
+    main()

python/ray/util/sgd/tests/test_torch.py

@@ -17,6 +17,7 @@ from ray.util.sgd.torch.constants import SCHEDULER_STEP
 from ray.util.sgd.utils import (check_for_failure, NUM_SAMPLES, BATCH_COUNT,
                                 BATCH_SIZE)
 
+from ray.util.sgd.data.examples import mlp_identity
 from ray.util.sgd.torch.examples.train_example import (
     model_creator, optimizer_creator, data_creator, LinearDataset)
 
@@ -32,6 +33,17 @@ def ray_start_2_cpus():
         dist.destroy_process_group()
 
 
+@pytest.fixture
+def ray_start_4_cpus():
+    address_info = ray.init(num_cpus=4)
+    yield address_info
+    # The code after the yield will run as teardown code.
+    ray.shutdown()
+    # Ensure that tests don't ALL fail
+    if dist.is_initialized():
+        dist.destroy_process_group()
+
+
 def test_single_step(ray_start_2_cpus):  # noqa: F811
     trainer = TorchTrainer(
         model_creator=model_creator,
@@ -120,12 +132,12 @@ def test_train(ray_start_2_cpus, num_workers):  # noqa: F811
 
 @pytest.mark.parametrize("num_workers", [1, 2] if dist.is_available() else [1])
 def test_multi_model(ray_start_2_cpus, num_workers):
-    def train(*, model=None, criterion=None, optimizer=None, dataloader=None):
+    def train(*, model=None, criterion=None, optimizer=None, iterator=None):
         model.train()
         train_loss = 0
         correct = 0
         total = 0
-        for batch_idx, (inputs, targets) in enumerate(dataloader):
+        for batch_idx, (inputs, targets) in enumerate(iterator):
             optimizer.zero_grad()
             outputs = model(inputs)
             loss = criterion(outputs, targets)
@@ -143,13 +155,14 @@ def test_multi_model(ray_start_2_cpus, num_workers):
 
     def train_epoch(self, iterator, info):
         result = {}
+        data = list(iterator)
         for i, (model, optimizer) in enumerate(
                 zip(self.models, self.optimizers)):
             result["model_{}".format(i)] = train(
                 model=model,
                 criterion=self.criterion,
                 optimizer=optimizer,
-                dataloader=iterator)
+                iterator=iter(data))
         return result
 
     def multi_model_creator(config):
@@ -310,6 +323,35 @@ def test_profiling(ray_start_2_cpus):  # noqa: F811
     trainer.shutdown()
 
 
+def test_dataset(ray_start_4_cpus):
+    """
+    This test tries training the mlp_identity example. We check the accuracy of
+    the model as an all inclusive way of ensuring that we are properly sharding
+    and iterating over the entire dataset (instead of repeating the first set
+    of points for example).
+    """
+
+    model_creator = mlp_identity.model_creator
+    optimizer_creator = mlp_identity.optimizer_creator
+    dataset_creator = mlp_identity.dataset_creator
+    trainer = TorchTrainer(
+        model_creator=model_creator,
+        data_creator=None,
+        optimizer_creator=optimizer_creator,
+        loss_creator=torch.nn.MSELoss,
+        num_workers=2,
+    )
+
+    dataset = dataset_creator()
+    for i in range(5):
+        trainer.train(dataset=dataset, num_steps=100)
+
+    input = mlp_identity.to_mat(0.5)
+    prediction = float(trainer.get_model()(input)[0][0])
+    assert 0.4 <= prediction <= 0.6
+    trainer.shutdown()
+
+
 def test_split_batch(ray_start_2_cpus):
     if not dist.is_available():
         return
@@ -591,20 +633,18 @@ def test_wrap_ddp(ray_start_2_cpus, tmp_path):  # noqa: F811
     trainer2.shutdown()
 
 
-def test_fail_with_recover(ray_start_2_cpus):  # noqa: F811
-    if not dist.is_available():
-        return
-
-    def single_loader(config):
-        dataset = LinearDataset(2, 5, size=1000000)
-        return DataLoader(dataset, batch_size=config.get("batch_size", 32))
-
-    def step_with_fail(self, **params):
+def gen_step_with_fail(num_fails):
+    def step_with_fail(self,
+                       num_steps=None,
+                       profile=False,
+                       info=None,
+                       dataset=None):
+        params = dict(num_steps=num_steps, profile=profile, info=info)
         remote_worker_stats = [
             w.train_epoch.remote(**params) for w in self.remote_workers
         ]
 
-        if self._num_failures < 3:
+        if self._num_failures < num_fails:
             time.sleep(1)  # Make the batch will fail correctly.
             ray.kill(self.remote_workers[0])
 
@@ -619,6 +659,19 @@ def test_fail_with_recover(ray_start_2_cpus):  # noqa: F811
 
         return success, None
 
+    return step_with_fail
+
+
+def test_fail_with_recover(ray_start_2_cpus):  # noqa: F811
+    if not dist.is_available():
+        return
+
+    def single_loader(config):
+        dataset = LinearDataset(2, 5, size=1000000)
+        return DataLoader(dataset, batch_size=config.get("batch_size", 32))
+
+    step_with_fail = gen_step_with_fail(3)
+
     with patch.object(TorchTrainer, "_train_epoch", step_with_fail):
         trainer1 = TorchTrainer(
             model_creator=model_creator,
@@ -642,25 +695,7 @@ def test_resize(ray_start_2_cpus):  # noqa: F811
         dataset = LinearDataset(2, 5, size=1000000)
         return DataLoader(dataset, batch_size=config.get("batch_size", 32))
 
-    def step_with_fail(self, **params):
-        remote_worker_stats = [
-            w.train_epoch.remote(**params) for w in self.remote_workers
-        ]
-
-        if self._num_failures < 1:
-            time.sleep(1)  # Make the batch will fail correctly.
-            ray.kill(self.remote_workers[0])
-
-        try:
-            local_worker_stats = self.local_worker.train_epoch(**params)
-        except RuntimeError:
-            return False, None
-
-        success = check_for_failure(remote_worker_stats)
-        if success:
-            return success, [local_worker_stats] + ray.get(remote_worker_stats)
-
-        return success, None
+    step_with_fail = gen_step_with_fail(1)
 
     with patch.object(TorchTrainer, "_train_epoch", step_with_fail):
         trainer1 = TorchTrainer(
@@ -691,25 +726,7 @@ def test_fail_twice(ray_start_2_cpus):  # noqa: F811
         dataset = LinearDataset(2, 5, size=1000000)
         return DataLoader(dataset, batch_size=config.get("batch_size", 32))
 
-    def step_with_fail(self, **params):
-        remote_worker_stats = [
-            w.train_epoch.remote(**params) for w in self.remote_workers
-        ]
-
-        if self._num_failures < 2:
-            time.sleep(1)  # Make the batch will fail correctly.
-            ray.kill(self.remote_workers[0])
-
-        try:
-            local_worker_stats = self.local_worker.train_epoch(**params)
-        except RuntimeError:
-            return False, None
-
-        success = check_for_failure(remote_worker_stats)
-        if success:
-            return success, [local_worker_stats] + ray.get(remote_worker_stats)
-
-        return success, None
+    step_with_fail = gen_step_with_fail(2)
 
     with patch.object(TorchTrainer, "_train_epoch", step_with_fail):
         trainer1 = TorchTrainer(

python/ray/util/sgd/torch/torch_runner.py

@@ -138,7 +138,9 @@ class TorchRunner:
     def setup_components(self):
         """Runs the creator functions without any distributed coordination."""
         logger.debug("Loading data.")
-        self._initialize_dataloaders()
+        if self.data_creator and callable(self.data_creator):
+            self._initialize_dataloaders()
+
         logger.debug("Creating model")
         self.models = self.model_creator(self.config)
         if not isinstance(self.models, Iterable):
@@ -181,7 +183,11 @@ class TorchRunner:
         """Finds a free port on the current node."""
         return utils.find_free_port()
 
-    def train_epoch(self, num_steps=None, profile=False, info=None):
+    def train_epoch(self,
+                    num_steps=None,
+                    profile=False,
+                    info=None,
+                    iterator=None):
         """Runs a training epoch and updates the model parameters."""
         logger.debug("Begin Training Step {}".format(self.epochs + 1))
         info = info or {}
@@ -193,9 +199,18 @@ class TorchRunner:
             SCHEDULER_STEP: self.scheduler_step_freq
         })
         with self.timers.record("train_epoch"):
-            iterator = self.train_loader
+            if iterator is None:
+                iterator = iter(self.train_loader)
+            else:
+                # Dataset will provide us with a list of tuples but we
+                # need two lists.
+                def format_batch(batch):
+                    features, targets = zip(*batch)
+                    return torch.cat(features), torch.cat(targets)
+
+                iterator = map(format_batch, iterator)
             if num_steps:
-                iterator = itertools.islice(iter(self.train_loader), num_steps)
+                iterator = itertools.islice(iterator, num_steps)
             train_stats = self.training_operator.train_epoch(iterator, info)
 
         self.epochs += 1

python/ray/util/sgd/torch/torch_trainer.py

@@ -17,6 +17,7 @@ from ray.util.sgd.torch.distributed_torch_runner import (
 from ray.util.sgd.utils import check_for_failure, NUM_SAMPLES, BATCH_SIZE
 from ray.util.sgd.torch.torch_runner import TorchRunner
 from ray.util.sgd.torch.constants import VALID_SCHEDULER_STEP
+from ray.util.sgd.data import Dataset
 
 logger = logging.getLogger(__name__)
 RESIZE_COOLDOWN_S = 10
@@ -194,11 +195,9 @@ class TorchTrainer:
                  "For more information, see "
                  "https://github.com/pytorch/examples/issues/467."))
 
-        if not (callable(model_creator) and callable(optimizer_creator)
-                and callable(data_creator)):
+        if not (callable(model_creator) and callable(optimizer_creator)):
             raise ValueError(
-                "Must provide a callable model_creator, optimizer_creator, "
-                "and data_creator.")
+                "Must provide a callable model_creator and optimizer_creator.")
 
         if num_replicas is not None:
             raise DeprecationWarning(
@@ -379,7 +378,8 @@ class TorchTrainer:
               profile=False,
               reduce_results=True,
               max_retries=3,
-              info=None):
+              info=None,
+              dataset=None):
         """Runs a training epoch.
 
         Calls `operator.train_epoch()` on N parallel workers simultaneously
@@ -405,6 +405,8 @@ class TorchTrainer:
                 in case of shared cluster usage. Defaults to 3.
             info (dict): Optional dictionary passed to the training
                 operator for ``train_epoch`` and ``train_batch``.
+            dataset (Dataset): Optional dataset to train with. If specified,
+                the dataloader passed in via data_creator will be ignored.
 
         Returns:
             (dict | list) A dictionary of metrics for training.
@@ -414,11 +416,14 @@ class TorchTrainer:
                 length will be equal to ``num_workers``.
         """
         assert max_retries >= 0, "`max_retries` must be non-negative."
+        assert isinstance(dataset, Dataset) is not None \
+            or self.data_creator, \
+            "Must specify either a data creator or a dataset"
         if self._should_resize():
             logger.info("Resize opportunity detected. Attempting to scale up.")
             self._resize_workers()
         success, worker_stats = self._train_epoch(
-            num_steps=num_steps, profile=profile, info=info)
+            num_steps=num_steps, profile=profile, info=info, dataset=dataset)
         # Fault handling
         for i in range(max_retries):
             if success:
@@ -429,7 +434,10 @@ class TorchTrainer:
             logger.info("Retrying training step with %d workers." %
                         (len(self.remote_workers) + 1))
             success, worker_stats = self._train_epoch(
-                num_steps=num_steps, profile=profile, info=info)
+                num_steps=num_steps,
+                profile=profile,
+                info=info,
+                dataset=dataset)
         if not success:
             raise RuntimeError("Training run failed.")
 
@@ -452,14 +460,26 @@ class TorchTrainer:
                 stats[stat_key] = worker_stats[0][stat_key]
         return stats
 
-    def _train_epoch(self, num_steps=None, profile=False, info=None):
+    def _train_epoch(self,
+                     num_steps=None,
+                     profile=False,
+                     info=None,
+                     dataset=None):
         params = dict(num_steps=num_steps, profile=profile, info=info)
-
-        remote_worker_stats = [
-            w.train_epoch.remote(**params) for w in self.remote_workers
-        ]
+        remote_worker_stats = []
+        if dataset:
+            dataset.set_num_shards(self.max_replicas)
+        for i, w in enumerate(self.remote_workers):
+            params = dict(num_steps=num_steps, profile=profile, info=info)
+            if dataset:
+                params["iterator"] = dataset.get_shard(i)
+            stats = w.train_epoch.remote(**params)
+            remote_worker_stats.append(stats)
 
         try:
+            if dataset:
+                params["iterator"] = dataset.get_shard(
+                    len(self.remote_workers))
             local_worker_stats = self.local_worker.train_epoch(**params)
         except RuntimeError as err:
             if "gloo" in err.args[0] and "Timed out" in err.args[0]: