Add dataset creation (#18940)

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: thomas <[email protected]>
(cherry picked from commit faa64c5)

Author: 3 people

requirements/app/app.txt

@@ -1,4 +1,4 @@
-lightning-cloud == 0.5.48  # Must be pinned to ensure compatibility
+lightning-cloud == 0.5.50  # Must be pinned to ensure compatibility
 packaging
 typing-extensions >=4.0.0, <4.8.0
 deepdiff >=5.7.0, <6.6.0

src/lightning/data/streaming/cache.py

@@ -19,7 +19,7 @@
 from lightning.data.datasets.env import _DistributedEnv
 from lightning.data.streaming.constants import (
     _INDEX_FILENAME,
-    _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_48,
+    _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_50,
     _TORCH_GREATER_EQUAL_2_1_0,
 )
 from lightning.data.streaming.item_loader import BaseItemLoader
@@ -29,7 +29,7 @@
 
 logger = logging.Logger(__name__)
 
-if _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_48:
+if _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_50:
     from lightning_cloud.resolver import _resolve_dir
 
 
@@ -67,8 +67,8 @@ def __init__(
         if not _TORCH_GREATER_EQUAL_2_1_0:
             raise ModuleNotFoundError("PyTorch version 2.1 or higher is required to use the cache.")
 
-        if not _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_48:
-            raise ModuleNotFoundError("Lightning Cloud 0.5.48 or higher is required to use the cache.")
+        if not _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_50:
+            raise ModuleNotFoundError("Lightning Cloud 0.5.50 or higher is required to use the cache.")
 
         input_dir = _resolve_dir(input_dir)
         self._cache_dir = input_dir.path

src/lightning/data/streaming/config.py

@@ -21,7 +21,7 @@
 from lightning.data.streaming.sampler import ChunkedIndex
 
 if _TORCH_GREATER_EQUAL_2_1_0:
-    from torch.utils._pytree import treespec_loads
+    from torch.utils._pytree import tree_unflatten, treespec_loads
 
 
 class ChunksConfig:
@@ -83,6 +83,24 @@ def data_format(self) -> Any:
             raise RuntimeError("The config should be defined.")
         return self._config["data_format"]
 
+    @property
+    def data_format_unflattened(self) -> Any:
+        if self._config is None:
+            raise RuntimeError("The config should be defined.")
+        return tree_unflatten(self._config["data_format"], self._config["data_spec"])
+
+    @property
+    def compression(self) -> Any:
+        if self._config is None:
+            raise RuntimeError("The config should be defined.")
+        return self._config["compression"]
+
+    @property
+    def chunk_bytes(self) -> int:
+        if self._config is None:
+            raise RuntimeError("The config should be defined.")
+        return self._config["chunk_bytes"]
+
     @property
     def config(self) -> Dict[str, Any]:
         if self._config is None:

src/lightning/data/streaming/constants.py

@@ -21,7 +21,7 @@
 # This is required for full pytree serialization / deserialization support
 _TORCH_GREATER_EQUAL_2_1_0 = RequirementCache("torch>=2.1.0")
 _VIZ_TRACKER_AVAILABLE = RequirementCache("viztracer")
-_LIGHTNING_CLOUD_GREATER_EQUAL_0_5_48 = RequirementCache("lightning-cloud>=0.5.48")
+_LIGHTNING_CLOUD_GREATER_EQUAL_0_5_50 = RequirementCache("lightning-cloud>=0.5.50")
 _BOTO3_AVAILABLE = RequirementCache("boto3")
 
 # DON'T CHANGE ORDER

src/lightning/data/streaming/data_processor.py

@@ -1,10 +1,12 @@
+import json
 import logging
 import os
 import signal
 import tempfile
 import traceback
 import types
 from abc import abstractmethod
+from dataclasses import dataclass
 from multiprocessing import Process, Queue
 from queue import Empty
 from shutil import copyfile, rmtree
@@ -23,7 +25,7 @@
     _BOTO3_AVAILABLE,
     _DEFAULT_FAST_DEV_RUN_ITEMS,
     _INDEX_FILENAME,
-    _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_48,
+    _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_50,
     _TORCH_GREATER_EQUAL_2_1_0,
 )
 from lightning.fabric.accelerators.cuda import is_cuda_available
@@ -35,10 +37,12 @@
 from lightning.fabric.utilities.distributed import group as _group
 
 if _TORCH_GREATER_EQUAL_2_1_0:
-    from torch.utils._pytree import tree_flatten, tree_unflatten
+    from torch.utils._pytree import tree_flatten, tree_unflatten, treespec_loads
 
-if _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_48:
+if _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_50:
+    from lightning_cloud.openapi import V1DatasetType
     from lightning_cloud.resolver import _resolve_dir
+    from lightning_cloud.utils.dataset import _create_dataset
 
 
 if _BOTO3_AVAILABLE:
@@ -191,7 +195,7 @@ def _upload_fn(upload_queue: Queue, remove_queue: Queue, cache_dir: str, output_
                 )
             except Exception as e:
                 print(e)
-        if os.path.isdir(output_dir.path):
+        elif os.path.isdir(output_dir.path):
             copyfile(local_filepath, os.path.join(output_dir.path, os.path.basename(local_filepath)))
         else:
             raise ValueError(f"The provided {output_dir.path} isn't supported.")
@@ -506,6 +510,16 @@ def __init__(self, *args: Any, **kwargs: Any) -> None:
         Process.__init__(self)
 
 
+@dataclass
+class _Result:
+    size: Optional[int] = None
+    num_bytes: Optional[str] = None
+    data_format: Optional[str] = None
+    compression: Optional[str] = None
+    num_chunks: Optional[int] = None
+    num_bytes_per_chunk: Optional[List[int]] = None
+
+
 T = TypeVar("T")
 
 
@@ -545,8 +559,8 @@ def listdir(self, path: str) -> List[str]:
     def __init__(self) -> None:
         self._name: Optional[str] = None
 
-    def _done(self, delete_cached_files: bool, output_dir: Dir) -> None:
-        pass
+    def _done(self, size: int, delete_cached_files: bool, output_dir: Dir) -> _Result:
+        return _Result(size=size)
 
 
 class DataChunkRecipe(DataRecipe):
@@ -576,7 +590,7 @@ def prepare_structure(self, input_dir: Optional[str]) -> List[T]:
     def prepare_item(self, item_metadata: T) -> Any:  # type: ignore
         """The return of this `prepare_item` method is persisted in chunked binary files."""
 
-    def _done(self, delete_cached_files: bool, output_dir: Dir) -> None:
+    def _done(self, size: int, delete_cached_files: bool, output_dir: Dir) -> _Result:
         num_nodes = _get_num_nodes()
         cache_dir = _get_cache_dir()
 
@@ -589,6 +603,26 @@ def _done(self, delete_cached_files: bool, output_dir: Dir) -> None:
         merge_cache._merge_no_wait(node_rank if num_nodes > 1 else None)
         self._upload_index(output_dir, cache_dir, num_nodes, node_rank)
 
+        if num_nodes == node_rank + 1:
+            with open(os.path.join(cache_dir, _INDEX_FILENAME)) as f:
+                config = json.load(f)
+
+            size = sum([c["dim"] if c["dim"] is not None else c["chunk_size"] for c in config["chunks"]])
+            num_bytes = sum([c["chunk_bytes"] for c in config["chunks"]])
+            data_format = tree_unflatten(config["config"]["data_format"], treespec_loads(config["config"]["data_spec"]))
+
+            return _Result(
+                size=size,
+                num_bytes=num_bytes,
+                data_format=data_format,
+                compression=config["config"]["compression"],
+                num_chunks=len(config["chunks"]),
+                num_bytes_per_chunk=[c["chunk_size"] for c in config["chunks"]],
+            )
+        return _Result(
+            size=size,
+        )
+
     def _upload_index(self, output_dir: Dir, cache_dir: str, num_nodes: int, node_rank: Optional[int]) -> None:
         """This method upload the index file to the remote cloud directory."""
         if output_dir.path is None and output_dir.url is None:
@@ -764,13 +798,31 @@ def run(self, data_recipe: DataRecipe) -> None:
                     has_failed = True
                     break
 
+        num_nodes = _get_num_nodes()
         # TODO: Understand why it hangs.
-        if _get_num_nodes() == 1:
+        if num_nodes == 1:
             for w in self.workers:
                 w.join(0)
 
         print("Workers are finished.")
-        data_recipe._done(self.delete_cached_files, self.output_dir)
+        result = data_recipe._done(num_items, self.delete_cached_files, self.output_dir)
+
+        if num_nodes == _get_node_rank() + 1:
+            _create_dataset(
+                input_dir=self.input_dir.path,
+                storage_dir=self.output_dir.path,
+                dataset_type=V1DatasetType.CHUNKED
+                if isinstance(data_recipe, DataChunkRecipe)
+                else V1DatasetType.TRANSFORMED,
+                empty=False,
+                size=result.size,
+                num_bytes=result.num_bytes,
+                data_format=result.data_format,
+                compression=result.compression,
+                num_chunks=result.num_chunks,
+                num_bytes_per_chunk=result.num_bytes_per_chunk,
+            )
+
         print("Finished data processing!")
 
         # TODO: Understand why it is required to avoid long shutdown.

src/lightning/data/streaming/dataset.py

@@ -19,12 +19,12 @@
 
 from lightning.data.datasets.env import _DistributedEnv, _WorkerEnv
 from lightning.data.streaming import Cache
-from lightning.data.streaming.constants import _INDEX_FILENAME, _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_48
+from lightning.data.streaming.constants import _INDEX_FILENAME, _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_50
 from lightning.data.streaming.item_loader import BaseItemLoader
 from lightning.data.streaming.sampler import ChunkedIndex
 from lightning.data.streaming.shuffle import FullShuffle, NoShuffle, Shuffle
 
-if _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_48:
+if _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_50:
     from lightning_cloud.resolver import _resolve_dir
 
 

src/lightning/data/streaming/functions.py

@@ -18,10 +18,10 @@
 from types import FunctionType
 from typing import Any, Callable, Optional, Sequence, Union
 
-from lightning.data.streaming.constants import _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_48, _TORCH_GREATER_EQUAL_2_1_0
+from lightning.data.streaming.constants import _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_50, _TORCH_GREATER_EQUAL_2_1_0
 from lightning.data.streaming.data_processor import DataChunkRecipe, DataProcessor, DataTransformRecipe
 
-if _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_48:
+if _LIGHTNING_CLOUD_GREATER_EQUAL_0_5_50:
     from lightning_cloud.resolver import _assert_dir_has_index_file, _assert_dir_is_empty, _execute, _resolve_dir
 
 if _TORCH_GREATER_EQUAL_2_1_0:

tests/tests_data/streaming/test_data_processor.py

@@ -102,7 +102,7 @@ def copy_file(local_filepath, *args):
         called = True
         from shutil import copyfile
 
-        copyfile(local_filepath, os.path.join(remote_output_dir.path, os.path.basename(local_filepath)))
+        copyfile(local_filepath, os.path.join(remote_output_dir, os.path.basename(local_filepath)))
 
     s3_client.client.upload_file = copy_file
 
@@ -420,8 +420,14 @@ def _broadcast_object(self, obj: Any) -> Any:
 def test_data_processsor_distributed(fast_dev_run, delete_cached_files, tmpdir, monkeypatch):
     """This test ensures the data optimizer works in a fully distributed settings."""
 
+    seed_everything(42)
+
     monkeypatch.setattr(data_processor_module.os, "_exit", mock.MagicMock())
 
+    _create_dataset_mock = mock.MagicMock()
+
+    monkeypatch.setattr(data_processor_module, "_create_dataset", _create_dataset_mock)
+
     from PIL import Image
 
     input_dir = os.path.join(tmpdir, "dataset")
@@ -501,6 +507,21 @@ def test_data_processsor_distributed(fast_dev_run, delete_cached_files, tmpdir,
 
     assert sorted(os.listdir(remote_output_dir)) == expected
 
+    _create_dataset_mock.assert_called()
+
+    assert _create_dataset_mock._mock_mock_calls[0].kwargs == {
+        "input_dir": str(input_dir),
+        "storage_dir": str(remote_output_dir),
+        "dataset_type": "CHUNKED",
+        "empty": False,
+        "size": 30,
+        "num_bytes": 26657,
+        "data_format": "jpeg",
+        "compression": None,
+        "num_chunks": 16,
+        "num_bytes_per_chunk": [2, 2, 2, 1, 2, 2, 2, 2, 2, 2, 2, 1, 2, 2, 2, 2],
+    }
+
 
 class TextTokenizeRecipe(DataChunkRecipe):
     def prepare_structure(self, input_dir: str) -> List[Any]: