import time
import traceback
from dataclasses import dataclass
from typing import Dict, MutableMapping, Optional, Union
import torch
import torch.multiprocessing as mp
from graphlearn_torch.distributed.rpc import barrier, rpc_is_initialized
from torch.multiprocessing import Manager
from gigl.common.data.dataloaders import (
SerializedTFRecordInfo,
TFDatasetOptions,
TFRecordDataLoader,
)
from gigl.common.logger import Logger
from gigl.src.common.types.graph_data import EdgeType, NodeType
from gigl.types.graph import (
DEFAULT_HOMOGENEOUS_EDGE_TYPE,
DEFAULT_HOMOGENEOUS_NODE_TYPE,
LoadedGraphTensors,
)
from gigl.utils.share_memory import share_memory
_ID_FMT = "{entity}_ids"
_FEATURE_FMT = "{entity}_features"
_NODE_KEY = "node"
_EDGE_KEY = "edge"
_POSITIVE_LABEL_KEY = "positive_label"
_NEGATIVE_LABEL_KEY = "negative_label"
@dataclass(frozen=True)
def _data_loading_process(
tf_record_dataloader: TFRecordDataLoader,
output_dict: MutableMapping[
str, Union[torch.Tensor, Dict[Union[NodeType, EdgeType], torch.Tensor]]
],
error_dict: MutableMapping[str, str],
entity_type: str,
serialized_tf_record_info: Union[
SerializedTFRecordInfo,
Dict[Union[NodeType, EdgeType], SerializedTFRecordInfo],
],
rank: int,
tf_dataset_options: TFDatasetOptions = TFDatasetOptions(),
) -> None:
"""
Spawned multiprocessing.Process which loads homogeneous or heterogeneous information for a specific entity type [node, edge, positive_label, negative_label]
and moves to shared memory. Also logs timing information for duration of loading. If an exception is thrown, its traceback will be stored in
the error_dict "error" field, since exceptions for spawned processes won't properly be raised to the parent process.
Args:
tf_record_dataloader (TFRecordDataLoader): TFRecordDataloader used for loading tensors from serialized tfrecords
output_dict (MutableMapping[str, Union[torch.Tensor, Dict[Union[NodeType, EdgeType], torch.Tensor]]]):
Dictionary initialized by mp.Manager().dict() in which outputs of tensor loading will be written to
error_dict (MutableMapping[str, str]): Dictionary initialized by mp.Manager().dict() in which error of errors in current process will be written to
entity_type (str): Entity type to prefix ids, features, and error keys with when
writing to the output_dict and error_dict fields
serialized_tf_record_info (Union[SerializedTFRecordInfo, Dict[NodeType, SerializedTFRecordInfo], Dict[EdgeType, SerializedTFRecordInfo]]):
Serialized information for current entity
rank (int): Rank of the current machine
tf_dataset_options (TFDatasetOptions): The options to use when building the dataset.
"""
# We add a try - except clause here to ensure that exceptions are properly circulated back to the parent process
try:
# To simplify the logic to proceed on a singular code path, we convert homogeneous inputs to heterogeneous just within the scope of this function
if isinstance(serialized_tf_record_info, SerializedTFRecordInfo):
serialized_tf_record_info = (
{DEFAULT_HOMOGENEOUS_NODE_TYPE: serialized_tf_record_info}
if serialized_tf_record_info.is_node_entity
else {DEFAULT_HOMOGENEOUS_EDGE_TYPE: serialized_tf_record_info}
)
is_input_homogeneous = True
else:
is_input_homogeneous = False
all_tf_record_uris = [
serialized_entity.tfrecord_uri_prefix.uri
for serialized_entity in serialized_tf_record_info.values()
]
start_time = time.time()
logger.info(
f"Rank {rank} has begun to load data from tfrecord directories: {all_tf_record_uris}"
)
ids: Dict[Union[NodeType, EdgeType], torch.Tensor] = {}
features: Dict[Union[NodeType, EdgeType], torch.Tensor] = {}
for (
graph_type,
serialized_entity_tf_record_info,
) in serialized_tf_record_info.items():
(
entity_ids,
entity_features,
) = tf_record_dataloader.load_as_torch_tensors(
serialized_tf_record_info=serialized_entity_tf_record_info,
tf_dataset_options=tf_dataset_options,
)
ids[graph_type] = entity_ids
logger.info(
f"Rank {rank} finished loading {entity_type} ids of shape {entity_ids.shape} for graph type {graph_type} from {serialized_entity_tf_record_info.tfrecord_uri_prefix.uri}"
)
if entity_features is not None:
features[graph_type] = entity_features
logger.info(
f"Rank {rank} finished loading {entity_type} features of shape {entity_features.shape} for graph type {graph_type} from {serialized_entity_tf_record_info.tfrecord_uri_prefix.uri}"
)
else:
logger.info(
f"Rank {rank} did not detect {entity_type} features for graph type {graph_type} from {serialized_entity_tf_record_info.tfrecord_uri_prefix.uri}"
)
logger.info(
f"Rank {rank} is attempting to share {entity_type} id memory for tfrecord directories: {all_tf_record_uris}"
)
share_memory(ids)
# We convert the ids back to homogeneous from the default heterogeneous setup if our provided input was homogeneous
if features:
logger.info(
f"Rank {rank} is attempting to share {entity_type} feature memory for tfrecord directories: {all_tf_record_uris}"
)
share_memory(features)
# We convert the features back to homogeneous from the default heterogeneous setup if our provided input was homogeneous
output_dict[_ID_FMT.format(entity=entity_type)] = (
list(ids.values())[0] if is_input_homogeneous else ids
)
if features:
output_dict[_FEATURE_FMT.format(entity=entity_type)] = (
list(features.values())[0] if is_input_homogeneous else features
)
logger.info(
f"Rank {rank} has finished loading {entity_type} data from tfrecord directories: {all_tf_record_uris}, elapsed time: {time.time() - start_time:.2f} seconds"
)
except Exception:
error_dict[entity_type] = traceback.format_exc()
[docs]
def load_torch_tensors_from_tf_record(
tf_record_dataloader: TFRecordDataLoader,
serialized_graph_metadata: SerializedGraphMetadata,
should_load_tensors_in_parallel: bool,
rank: int = 0,
node_tf_dataset_options: TFDatasetOptions = TFDatasetOptions(),
edge_tf_dataset_options: TFDatasetOptions = TFDatasetOptions(),
) -> LoadedGraphTensors:
"""
Loads all torch tensors from a SerializedGraphMetadata object for all entity [node, edge, positive_label, negative_label] and edge / node types.
Running these processes in parallel slows the runtime of each individual process, but may still result in a net speedup across all entity types. As a result,
there is a tradeoff that needs to be made between parallel and sequential tensor loading, which is why we don't parallelize across node and edge types. We enable
the `should_load_tensors_in_parallel` to allow some customization for loading strategies based on the input data.
Args:
tf_record_dataloader (TFRecordDataLoader): TFRecordDataloader used for loading tensors from serialized tfrecords
serialized_graph_metadata (SerializedGraphMetadata): Serialized graph metadata contained serialized information for loading tfrecords across node and edge types
should_load_tensors_in_parallel (bool): Whether tensors should be loaded from serialized information in parallel or in sequence across the [node, edge, pos_label, neg_label] entity types.
rank (int): Rank on current machine
node_tf_dataset_options (TFDatasetOptions): The options to use for nodes when building the dataset.
edge_tf_dataset_options (TFDatasetOptions): The options to use for edges when building the dataset.
Returns:
loaded_graph_tensors (LoadedGraphTensors): Unpartitioned Graph Tensors
"""
logger.info(f"Rank {rank} starting loading torch tensors from serialized info ...")
start_time = time.time()
manager = Manager()
# By default, torch processes are created using the `fork` method, which makes a copy of the entire process. This can be problematic in multi-threaded settings,
# especially when working with TensorFlow, since this includes all threads, which can lead to deadlocks or other synchronization issues. As a result, we set the
# start method to spawn, which creates a new Python interpreter process and is much safer with multi-threading applications.
ctx = mp.get_context("spawn")
node_output_dict: MutableMapping[
str, Union[torch.Tensor, Dict[NodeType, torch.Tensor]]
] = manager.dict()
edge_output_dict: MutableMapping[
str, Union[torch.Tensor, Dict[EdgeType, torch.Tensor]]
] = manager.dict()
error_dict: MutableMapping[str, str] = manager.dict()
node_data_loading_process = ctx.Process(
target=_data_loading_process,
kwargs={
"tf_record_dataloader": tf_record_dataloader,
"output_dict": node_output_dict,
"error_dict": error_dict,
"entity_type": _NODE_KEY,
"serialized_tf_record_info": serialized_graph_metadata.node_entity_info,
"rank": rank,
"tf_dataset_options": node_tf_dataset_options,
},
)
edge_data_loading_process = ctx.Process(
target=_data_loading_process,
kwargs={
"tf_record_dataloader": tf_record_dataloader,
"output_dict": edge_output_dict,
"error_dict": error_dict,
"entity_type": _EDGE_KEY,
"serialized_tf_record_info": serialized_graph_metadata.edge_entity_info,
"rank": rank,
"tf_dataset_options": edge_tf_dataset_options,
},
)
if serialized_graph_metadata.positive_label_entity_info is not None:
positive_label_data_loading_process = ctx.Process(
target=_data_loading_process,
kwargs={
"tf_record_dataloader": tf_record_dataloader,
"output_dict": edge_output_dict,
"error_dict": error_dict,
"entity_type": _POSITIVE_LABEL_KEY,
"serialized_tf_record_info": serialized_graph_metadata.positive_label_entity_info,
"rank": rank,
},
)
else:
logger.info(f"No positive labels detected from input data")
if serialized_graph_metadata.negative_label_entity_info is not None:
negative_label_data_loading_process = ctx.Process(
target=_data_loading_process,
kwargs={
"tf_record_dataloader": tf_record_dataloader,
"output_dict": edge_output_dict,
"error_dict": error_dict,
"entity_type": _NEGATIVE_LABEL_KEY,
"serialized_tf_record_info": serialized_graph_metadata.negative_label_entity_info,
"rank": rank,
},
)
else:
logger.info(f"No negative labels detected from input data")
if should_load_tensors_in_parallel:
# In this setting, we start all the processes at once and join them at the end to achieve parallelized tensor loading
logger.info("Loading Serialized TFRecord Data in Parallel ...")
node_data_loading_process.start()
edge_data_loading_process.start()
if serialized_graph_metadata.positive_label_entity_info is not None:
positive_label_data_loading_process.start()
if serialized_graph_metadata.negative_label_entity_info is not None:
negative_label_data_loading_process.start()
node_data_loading_process.join()
edge_data_loading_process.join()
if serialized_graph_metadata.positive_label_entity_info is not None:
positive_label_data_loading_process.join()
if serialized_graph_metadata.negative_label_entity_info is not None:
negative_label_data_loading_process.join()
else:
# In this setting, we start and join each process one-at-a-time in order to achieve sequential tensor loading
logger.info("Loading Serialized TFRecord Data in Sequence ...")
# Here we launch edge loading process first since experimentally
# we have found that, since edge data is larger than node data,
# launching edge before launching node reduces peak memory usage
# compared with first launching node loading and then launching edge loading.
edge_data_loading_process.start()
edge_data_loading_process.join()
node_data_loading_process.start()
node_data_loading_process.join()
if serialized_graph_metadata.positive_label_entity_info is not None:
positive_label_data_loading_process.start()
positive_label_data_loading_process.join()
if serialized_graph_metadata.negative_label_entity_info is not None:
negative_label_data_loading_process.start()
negative_label_data_loading_process.join()
if error_dict:
for entity_type, traceback in error_dict.items():
logger.error(
f"Identified error in {entity_type} data loading process: \n{traceback}"
)
raise ValueError(
f"Raised error in data loading processes for entity types {error_dict.keys()}."
)
node_ids = node_output_dict[_ID_FMT.format(entity=_NODE_KEY)]
node_features = node_output_dict.get(_FEATURE_FMT.format(entity=_NODE_KEY), None)
edge_index = edge_output_dict[_ID_FMT.format(entity=_EDGE_KEY)]
edge_features = edge_output_dict.get(_FEATURE_FMT.format(entity=_EDGE_KEY), None)
positive_labels = edge_output_dict.get(
_ID_FMT.format(entity=_POSITIVE_LABEL_KEY), None
)
negative_labels = edge_output_dict.get(
_ID_FMT.format(entity=_NEGATIVE_LABEL_KEY), None
)
if rpc_is_initialized():
logger.info(
f"Rank {rank} has finished loading data in {time.time() - start_time:.2f} seconds. Wait for other ranks to finish loading data from tfrecords"
)
barrier()
logger.info(
f"All ranks have finished loading data from tfrecords, rank {rank} finished in {time.time() - start_time:.2f} seconds"
)
return LoadedGraphTensors(
node_ids=node_ids,
node_features=node_features,
edge_index=edge_index,
edge_features=edge_features,
positive_label=positive_labels,
negative_label=negative_labels,
)
