Skip to content

bionemo-size-aware-batching

To install, execute the following: 

pip install -e .

To run unit tests, execute: 

pytest -v .

Summary of Usage

This package provides a simple way to create mini-batches in a memory consumption-aware (or size-aware) manner, making it useful for tasks like training models on datasets with varying memory requirements. The usage typically consists of the following steps:

  1. Use the collect_cuda_peak_alloc function to collect CUDA peak memory allocation statistics for a user-defined workflow. It's expected that the user-defined workflow will return a list of features extracted from the data so that the memory model in the following step can use these features to predict the memory allocation.
  2. User defines and trains a memory model using the features and memory allocation data from previous step. This memory model can then be used to predict memory consumption.
  3. Use SizeAwareBatchSampler or size_aware_batching with the memory model prediction (from the previous step) to build batch of data so that the resulting mini-batches do not exceed a specified maximum total memory size.

In addition, this package provides one solution to create homogeneous mini-batches, which can be useful to reduce the padding when aligning the shape of inputs when training or evaluating the models. This BucketBatchSampler can be used in conjunction with torch.utils.data.BatchSampler, SizeAwareBatchSampler or other user-defined batch samplers. This usage can leverage the create_buckets function and follow the steps below:

  1. Gather the tensor sizes of elements in the dataset, which are the shapes of tensors in a specific dimension where you want to reduce the padding.
  2. Provide your own bucket boundaries based on the tensor sizes, or create bucket boundaries with create_buckets function with the tensor sizes and bucket maximal width and the minimal bucket count. The create_buckets function will try to create buckets with smallest widths and element counts >= minimal bucket count, unless the maximal width or the boundary is reached.
  3. Use BucketBatchSampler with base batch sampler like torch.utils.data.BatchSampler or SizeAwareBatchSampler for each bucket. The BucketBatchSampler will select a bucket each time and generate a mini-batch from this bucket using the base batch sampler for this bucket. As such, the padding necessary for the generated mini-batches will be always smaller than the width of buckets.

Refer to the later sections for the API documentation and examples on how to achieve each of the steps above.

utils Module

  • collect_cuda_peak_alloc: A function that collects CUDA peak memory allocation statistics and features to be used for memory usage prediction for a given workflow.

  • create_buckets: A function to create buckets for a list of integers with pre-defined maximal width of ranges and minimal bucket sizes.

sampler Module

  • size_aware_batching: A generator that batches elements from an iterable while ensuring that the total size of each batch does not exceed a specified maximum.
  • SizeAwareBatchSampler: A class that batches elements of varying sizes while ensuring that the total size of each batch does not exceed a specified maximum.
  • BucketBatchSampler: A class that groups elements of varying sizes based on predefined bucket ranges, and create batches with elements from each bucket to ensure that each batch has elements with homogeneous sizes.

API reference and examples

utils

collect_cuda_peak_alloc

def collect_cuda_peak_alloc(
    dataset: Iterable[Data],
    work: Callable[[Data], Feature],
    device: torch.device,
    cleanup: Optional[Callable[[], None]] = None
) -> Tuple[List[Feature], List[int]]

Collects CUDA peak memory allocation statistics for a given workflow.

This function iterates through the provided dataset, applies the given feature function to each data point, and records the peak CUDA memory allocation during this process. The features extracted from the data points are collected along with their corresponding memory usage statistics.

Note that the first few iterations of the workflow might result in smaller memory allocations due to uninitialized data (e.g., internal PyTorch buffers). Therefore, users may want to skip these initial data points when analyzing the results.

Arguments:

  • dataset - An iterable containing the input data.
  • work - A function that takes a data point and returns its corresponding feature. This is where the main computation happens and memory allocations are tracked.
  • device - The target Torch CUDA device.
  • cleanup - A function that is called after each iteration to perform any necessary cleanup.

Returns:

A tuple containing the collected features and their corresponding memory usage statistics.

Raises:

  • ValueError - If the provided device is not a CUDA device.

Examples:

>>> import torch
>>> from bionemo.size_aware_batching.utils import collect_cuda_peak_alloc


>>> # prepare dataset, model and other components of a workflow
>>> # for which the user want to collect CUDA peak memory allocation statistics
>>> dataset, model, optimizer = ...
>>> # Set the target Torch CUDA device.
>>> device = torch.device("cuda:0")
>>> model = model.to(device)

>>> # Define a function that takes an element of the dataset as input and
>>> # do a training step
>>> def work(data):
...     # example body of a training loop
...     optimizer.zero_grad()
...     output = model(data.to(device))
...     loss = compute_loss(output)
...     loss.backward()
...     optimizer.step()
...     # extract the feature for later to be modeled or analyzed
...     return featurize(data)

>>> # can optionally use a cleanup function to release the references
>>> # hold during the work(). This cleanup function will be called
>>> # at the end of each step before garbage collection and memory allocations measurement
>>> def cleanup():
...     model.zero_grad(set_to_none=True)

>>> # Collect features (i.e., model outputs) and memory usage statistics for the workflow.
>>> features, alloc_peaks = collect_cuda_peak_alloc(
...     dataset=batches,
...     work=work,
...     device=device,
...     cleanup=cleanup,
... )


>>> # use features and alloc_peaks as needed, e.g., fit a model
>>> # that can use these statistics to predict memory usage
>>> memory_model = ...
>>> memory_model.fit(features, alloc_peaks)

create_buckets

def create_buckets(sizes: torch.Tensor, max_width: int,
                   min_bucket_count: int) -> Buckets

Create buckets for a list of integers with pre-defined maximal width of interval and minimal bucket count.

It will return a named tuple containing the bucket boundaries and the actual bucket sizes. e.g. torch.tensor([0, 5, 7]), torch.tensor([3,2]): specifies 2 buckets: one with range 0<= sizes < 5, width=5 and 3 elements and the other one with range 5 <= sizes < 7, width=2 and 2 elements.

Arguments:

  • sizes - An 1D tensor of integers.
  • max_width - The maximum width of a bucket, should be a positive integer.
  • min_bucket_count - The minimum count of a bucket, should be a positive integer. Bucket size may be smaller than min_bucket_count if its width reaches max_width.

Raises:

  • ValueError - If the provided sizes is empty, or not integers.
  • ValueError - If max_width is not a positive integer or min_bucket_count is not a positive integer.

Returns:

A named tuple containing bucket boundaries in ascending order and the number of elements in each bucket.

Examples:

>>> import torch
>>> from bionemo.size_aware_batching.utils import create_buckets

>>> sizes = torch.tensor([1, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 22, 22, 22, 22])
>>> buckets = create_buckets(sizes, max_width=5, min_bucket_count=10)
>>> # 5 buckets: 1 <= sizes < 6, 6 <= sizes < 11, 11 <= sizes < 16, 16 <= sizes < 21, 21 <= sizes < 23
>>> print(buckets.bucket_boundaries)
tensor([ 1,  6, 11, 16, 21, 23])

>>> # each with 12, 0, 0, 0, 4 elements respectively.
>>> print(buckets.bucket_sizes)
tensor([12,  0,  0,  0,  4])

>>> sizes = torch.arange(20)
>>> # min_bucket_count is used to control bucket size
>>> buckets = create_buckets(sizes, max_width=10, min_bucket_count=5)
>>> print(buckets.bucket_boundaries)
tensor([ 0,  5, 10, 15, 20])

>>> print(buckets.bucket_sizes)
tensor([5, 5, 5, 5])

sampler

size_aware_batching

def size_aware_batching(
    dataset: Iterable[Data],
    sizeof: Callable[[Data], Real],
    max_total_size: Real,
    collate_fn: Optional[Callable[[Iterable[Data]], BatchCollated]] = None,
    info_logger: Optional[Callable[[str], None]] = None,
    warn_logger: Optional[Callable[[str], None]] = None
) -> Iterator[Union[List[Data], BatchCollated]]

A generator that batches elements from an iterable while ensuring that the total size of each batch does not exceed a specified maximum. Here the size can be a measurement of memory consumption of the elements in the batch. This can be useful for both indexible data or non-indexible but iterable data.

Arguments:

  • dataset - The input iterable.
  • sizeof - A function or mapping that returns the "size" of each element in dataset. E.g., this can be used to determine how much memory an element consumes. Its return type must be comparable with max_total_size and it must be addable (operator +).
  • max_total_size - The maximum total "size" of each batch. The semantics of "size" is defined by the sizeof argument. The type of this value must be comparable with the return type of sizeof, i.e., the operator < and == must be meaningful.
  • collate_fn - An optional function to collate batches. Defaults to None, in which case each batch is a list of elements from the input dataset
  • info_logger - A function to log info. Defaults to None.
  • warn_logger - A function to log warnings. Defaults to None.

Yields:

A generator that yields batches from dataset.

Assumptions 1. Linear complexity. This function consumes the given Iterable of data (dataset) once, by going over the data item one by one to build a batch and yield it as soon as the addition of the next data item to the batch would exceed max_total_size or if the batch is the last one (end of iteration) 2. Additive size measurement. For the general usage case of building mini-batches with a threshold of the batch's memory consumption, it assumes that the size of the batch is the sum of all elements in the batch (additive property). 3. Comparable type of max_total_size and sizeof's return. sizeof's return values must be compared with max_total_size to threshold the size of batches

Caveat - 1 - The generated batch sizes may have large variance - how to workaround: filter the output of this generator using a batch size threshold - 2 - The number of batches may vary a lot across different epochs. - how to workaround: increase the number of steps that compose an epoch, e.g., in the Lightning training/validation loop, which effectively increases the input dataset size per epoch

Example

>>> import torch
>>> from torch.utils.data import default_collate
>>> from bionemo.size_aware_batching.sampler import size_aware_batching

>>> # Define a sample dataset with torch.tensor
>>> dataset = [torch.tensor([1, 2]), torch.tensor([3, 4]), torch.tensor([5, 6]),
...            torch.tensor([7, 8]), torch.tensor([9, 10])]

>>> # Define a sizeof function that returns the size of each tensor
>>> def sizeof(x):
...     return x.numel()

>>> # Create a generator with max_total_size=4 and default_collate_fn
>>> gen = size_aware_batching(dataset, sizeof, 4, collate_fn=default_collate)
>>> batches = list(gen)
>>> print(batches)
    [tensor([[1, 2], [3, 4]]), tensor([[5, 6], [7, 8]]), tensor([[9, 10]])]

SizeAwareBatchSampler Objects

class SizeAwareBatchSampler(Sampler[List[int]])

A sampler that batches elements of varying sizes while ensuring that the total size of each batch does not exceed a specified maximum.

This is useful when dealing with datasets where each element has a different size, such as graphs or sequences of varying lengths. The sampler uses a provided sizeof function to determine the size of each element in the dataset and ensures that the total size of each batch does not exceed the specified max_total_size.

Examples:

>>> import torch
>>> from bionemo.size_aware_batching.sampler import SizeAwareBatchSampler


>>> # Define a sample dataset with torch.tensor
>>> dataset = [torch.tensor([1, 2]), torch.tensor([3, 4]), torch.tensor([5, 6]),
...            torch.tensor([7, 8]), torch.tensor([9, 10])]


>>> # Define a function that returns the size of each element in the dataset.
>>> def sizeof(index):
...     return dataset[index].numel()


>>> # Create a SizeAwareBatchSampler with a maximum total batch size of 10.
>>> batch_sampler = SizeAwareBatchSampler(
...     sampler=torch.utils.data.SequentialSampler(dataset),
...     sizeof=sizeof,
...     max_total_size=4
... )


>>> # Iterate over batches of indices that do not exceed the maximum total size.
>>> print(list(batch_sampler))
    [[0, 1], [2, 3], [4]]

__init__

def __init__(sampler: Union[Sampler[List[int]], Iterable[int]],
             sizeof: Callable[[int], Real],
             max_total_size: Real,
             info_logger: Optional[Callable[[str], None]] = None,
             warn_logger: Optional[Callable[[str], None]] = None) -> None

Initializes the SizeAwareBatchSampler.

Arguments:

  • sampler - The underlying sampler.
  • sizeof - A function that returns the size at each index. E.g., this can used to determine how much memory an element consumes. Its return type must be comparable with max_total_size and it must be addable (operator +).
  • max_total_size - The maximum total size of a mini-batch. The semantics of "size" is defined by the sizeof argument. The type of this value must be comparable with the return type of sizeof, i.e., the operator < and == must be meaningful.
  • info_logger - A function to log info. Defaults to None.
  • warn_logger - A function to log warnings. Defaults None.

Raises:

  • TypeError - If sampler is not an instance of Sampler or Iterable, or if sizeof is not a callable, dictionary, or sequence container.
  • ValueError - If max_total_size is not a positive number.

__iter__

def __iter__() -> Iterator[List[int]]

Iterate over batches of indices.

This function yields batches of indices that do not exceed the maximum total size.

Yields:

A batch of indices that do not exceed the maximum total size.

BucketBatchSampler Objects

class BucketBatchSampler(Sampler[List[int]])

A batch sampler to create batches with sizes of elements from each pre-defined bucket ranges.

Elements of the dataset are first grouped into each bucket based on the bucket ranges and the sizes of elements. Then, a base batch sampler is used for each bucket to create mini-batches.

The bucket ranges are specified by bucket_boundaries, which will be first sorted internally and used to create len(bucket_boundaries) - 1 left-closed right-open intervals. e.g. if bucket_boundaries tensor is [10, 5, 0, 16], it will be sorted as [0, 5, 10, 16] and 3 buckets will be created with ranges: [0, 5), [5, 10), [10, 16).

The base batch sampler will be created by passing the element indices in each bucket as the data source, and base_batch_sampler_shared_kwargs and base_batch_sampler_individual_kwargs to the constructor of the base batch sampler class specified as base_batch_sampler_class. e.g. base_batch_sampler_shared_kwargs = {'drop_last': True} and base_batch_sampler_individual_kwargs = {'batch_size': [8,10,12]} will be used to create 3 batch samplers with drop_last=True and batch_size=8, 10 and 12, and initialized like base_batch_sampler_class(bucket_element_indices[0], batch_size=8, drop_last=True).

In the __iter__ method, if shuffle is True, the element indices in each bucket will be shuffled, and a bucket is randomly selected each time to create a mini-batch. If shuffle is False, there is no shuffle on element indices, and the bucket is selected in ascending order of its interval boundaries.

This class is used to create homogeneous batches of data for training or evaluation, and reduce the padding necessary to align the shape of elements.

Modified from https://github.com/rssrwn/semla-flow/blob/main/semlaflow/data/util.py

Examples:

>>> import torch
>>> from bionemo.size_aware_batching.sampler import BucketBatchSampler

>>> # Define the sizes for a dataset
>>> sizes = torch.arange(25)
>>> # Define bucket ranges
>>> bucket_boundaries = torch.tensor([0, 6, 15, 25])

>>> # Create a bucket batch sampler with torch.utils.data.BatchSampler as base batch sampler
>>> # As there are 3 buckets, there will be 3 base batch samplers with batch sizes 2, 3, and 5.
>>> batch_sampler = BucketBatchSampler(
        sizes=sizes,
        bucket_boundaries=bucket_boundaries,
        base_batch_sampler_class=torch.utils.data.BatchSampler,
        base_batch_sampler_shared_kwargs={'drop_last': False},
        base_batch_sampler_individual_kwargs={'batch_size': [2,3,5]},
        shuffle=False,
    )

>>> # Iterate over batches of indices that lies in the same bucket and with different batch sizes.
>>> print(list(batch_sampler))
[[0, 1], [2, 3], [4, 5], [6, 7, 8], [9, 10, 11], [12, 13, 14], [15, 16, 17, 18, 19], [20, 21, 22, 23, 24]]

>>> # randomize the dataset and buckets
>>> batch_sampler = BucketBatchSampler(
        sizes=sizes,
        bucket_boundaries=bucket_boundaries,
        base_batch_sampler_class=torch.utils.data.BatchSampler,
        base_batch_sampler_shared_kwargs={'drop_last': False},
        base_batch_sampler_individual_kwargs={'batch_size': [2,3,5]},
        shuffle=True,
        generator=torch.Generator().manual_seed(0),
    )
>>> print(list(batch_sampler))
[[24, 17, 16, 22, 19], [2, 5], [12, 10, 11], [3, 0], [15, 18, 20, 21, 23], [7, 13, 6], [14, 9, 8], [1, 4]]
>>> print(list(batch_sampler))
[[14, 9, 13], [23, 16, 20, 21, 15], [5, 0], [8, 10, 11], [17, 24, 22, 18, 19], [12, 6, 7], [4, 2], [3, 1]]

>>> # Combine with SizeAwareBatchSampler to control the cost of each batch
>>> from bionemo.size_aware_batching.sampler import SizeAwareBatchSampler
>>> item_costs = sizes.tolist()
>>> def cost_of_element(index):
        return item_costs[index]
>>> batch_sampler = BucketBatchSampler(
        sizes=sizes,
        bucket_boundaries=bucket_boundaries,
        base_batch_sampler_class=SizeAwareBatchSampler,
        base_batch_sampler_shared_kwargs={"sizeof": cost_of_element, "max_total_size": 40},
        base_batch_sampler_individual_kwargs={},
        shuffle=True,
        generator=torch.Generator().manual_seed(0),
    )
>>> print(list(iter(batch_sampler)))
[[24], [2, 5, 3, 0, 1, 4], [12, 10, 11, 7], [13, 6, 14], [17, 16], [22], [19, 15], [9, 8], [18, 20], [21], [23]]

__init__

def __init__(sizes: torch.Tensor,
             bucket_boundaries: torch.Tensor,
             base_batch_sampler_class: Type[Sampler],
             base_batch_sampler_shared_kwargs: Optional[Dict[str, Any]] = None,
             base_batch_sampler_individual_kwargs: Optional[Dict[
                 str, Iterable]] = None,
             shuffle: Optional[bool] = True,
             generator: Optional[torch.Generator] = None) -> None

Initializes the BucketBatchSampler.

Arguments:

  • sizes - A 1D tensor of real numbers representing the size of each element in the dataset.
  • bucket_boundaries - A 1D tensor of real numbers representing the boundaries of the bucket ranges. It will be first sorted and used to create len(bucket_boundaries) - 1 left-closed right-open intervals as bucket ranges. It should not contain any duplicate values.
  • base_batch_sampler_class - Base batch sampler class type, which will be used for each bucket, and initialized with the bucket element indices, base_batch_sampler_shared_kwargs and the corresponding base_batch_sampler_individual_kwargs.
  • base_batch_sampler_shared_kwargs - Shared keyword argument dictionary used to initialize all base batch samplers for all buckets. Sufficient and valid arguments should be provided for base_batch_sampler_class with base_batch_sampler_individual_kwargs. Default to {}.
  • base_batch_sampler_individual_kwargs - Keyword argument dictionary used to initialize each bucket batch sampler with the corresponding key value pairs. Length of each value in this dict must be equal to len(bucket_boundaries) - 1 (the number of buckets). Sufficient and valid arguments should be provided for base_batch_sampler_class with base_batch_sampler_shared_kwargs. Default to {}.
  • shuffle - A boolean indicating whether to shuffle the dataset and buckets. Defaults to True.
  • generator - Generator used in sampling. Defaults to None.

Raises:

  • ValueError - If sizes is not a 1D tensor of real numbers.
  • ValueError - If bucket_boundaries is not a 1D tensor of real numbers.
  • ValueError - If base_batch_sampler_individual_kwargs or base_batch_sampler_individual_kwargs is not a keyword argument dictionary.
  • ValueError - If the length of values in the dict of base_batch_sampler_individual_kwargs must be equal to len(bucket_boundaries) - 1.
  • RuntimeError - If there is no elements with sizes inside the ranges specified by bucket_boundaries.

__len__

def __len__() -> int

Get the number of batches.

Can only be called if the base_batch_sampler_class has len() implemented

Returns:

  • int - Number of batches

__iter__

def __iter__() -> Iterator[List[int]]

Iterate over batches of indices.

This function yields batches of indices of elements with sizes from each bucket range.

Yields:

  • List[int] - A batch of indices of elements with sizes from each bucket range.