Datasets and data loaders

This tutorial shows how to define Dataset and DataLoader compatible with PyTorch and containing metatensor data (i.e. data stored in metatensor.torch.TensorMap) in addition to more usual types of data.

import os

import torch

from metatensor.learn.data import DataLoader, Dataset
from metatensor.torch import Labels, TensorBlock, TensorMap

Let’s define a simple dummy dataset with two fields, named ‘x’ and ‘y’. Every field in the Dataset must be a list of objects corresponding to the different samples in this dataset.

Let’s define our x data as a list of random tensors, and our y data as a list of integers enumerating the samples.

n_samples = 5
x_data = [torch.randn(3) for _ in range(n_samples)]
y_data = [i for i in range(n_samples)]

In-memory dataset

We are ready to build out first dataset. The simplest use case is when all data is in memory. In this case, we can pass the data directly to the Dataset constructor as keyword arguments, named and ordered according to how we want the data to be returned when we access samples in the dataset.

in_memory_dataset = Dataset(x=x_data, y=y_data)

We can now access samples in the dataset. The returned object is a named tuple with fields corresponding to the keyword arguments given to the :py:class:Dataset` constructor (here ``x and y).

print(in_memory_dataset[0])

Sample(x=tensor([-1.6455,  0.9555, -0.5801]), y=0)

One can also iterate over the samples in the dataset as follows:

for sample in in_memory_dataset:
    print(sample)

Sample(x=tensor([-1.6455,  0.9555, -0.5801]), y=0)
Sample(x=tensor([-0.1646, -1.3758,  1.3614]), y=1)
Sample(x=tensor([ 1.5761, -0.2646, -0.1489]), y=2)
Sample(x=tensor([0.7162, 0.2067, 1.0703]), y=3)
Sample(x=tensor([-0.5449, -0.5098,  0.0966]), y=4)

Any number of named data fields can be passed to the Dataset constructor, as long as they are all uniquely named, and are all lists of the same length. The elements of each list can be any type of object (integer, string, torch Tensor, etc.), as long as it is the type same for all samples in the respective field.

For example, here we are creating a dataset of torch tensors (x), integers (y), and strings (z).

bigger_dataset = Dataset(x=x_data, y=y_data, z=["a", "b", "c", "d", "e"])
print(bigger_dataset[0])
print("Sample 4, z field:", bigger_dataset[4].z)

Sample(x=tensor([-1.6455,  0.9555, -0.5801]), y=0, z='a')
Sample 4, z field: e

Mixed in-memory / on-disk dataset

Now suppose we have a large dataset, where the x data is too large to fit in memory. In this case, we might want to lazily load data when training a model with minibatches.

Let’s save the x data to disk to simulate this use case.

# Create a directory to save the dummy x data to disk
os.makedirs("data", exist_ok=True)

for i, x in enumerate(x_data):
    torch.save(x, f"data/x_{i}.pt")

In order for the x data to be loaded lazily, we need to give the Dataset a load function that loads a single sample into memory. This can a function of arbitrary complexity, taking a single argument which is the numeric index (between 0 and len(dataset)) of the sample to load

def load_x(sample_id):
    """
    Loads the x data for the sample indexed by `sample_id` from disk and returns the
    object in memory
    """
    print(f"loading x for sample {sample_id}")
    return torch.load(f"data/x_{sample_id}.pt")


print("load_x called with sample index 0:", load_x(0))

loading x for sample 0
/home/runner/work/metatensor/metatensor/python/examples/learn/1-dataset-dataloader.py:107: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  return torch.load(f"data/x_{sample_id}.pt")
load_x called with sample index 0: tensor([-1.6455,  0.9555, -0.5801])

Now when we define a dataset, the ‘x’ data field can be passed as a callable.

mixed_dataset = Dataset(x=load_x, y=y_data)
print(mixed_dataset[3])

loading x for sample 3
Sample(x=tensor([0.7162, 0.2067, 1.0703]), y=3)

On-disk dataset

Finally, suppose we have a large dataset, where both the x and y data are too large to fit in memory. In this case, we might want to lazily load all data when training a model with minibatches.

Let’s save the y data to disk as well to simulate this use case.

for i, y in enumerate(y_data):
    torch.save(y, f"data/y_{i}.pt")


def load_y(sample_id):
    """
    Loads the y data for the sample indexed by `sample_id` from disk and
    returns the object in memory
    """
    print(f"loading y for sample {sample_id}")
    return torch.load(f"data/y_{sample_id}.pt")


print("load_y called with sample index 0:", load_y(0))

loading y for sample 0
/home/runner/work/metatensor/metatensor/python/examples/learn/1-dataset-dataloader.py:140: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  return torch.load(f"data/y_{sample_id}.pt")
load_y called with sample index 0: 0

Now when we define a dataset, as all the fields are to be lazily loaded, we need to indicate how many samples are in the dataset with the size argument.

Internally, the Dataset class infers the unique sample indexes as a continuous integer sequence starting from 0 to size - 1 (inclusive). In this case, sample indexes are therefore [0, 1, 2, 3, 4]. These indexes are used to lazily load the data upon access.

on_disk_dataset = Dataset(x=load_x, y=load_y, size=n_samples)
print(on_disk_dataset[2])

loading x for sample 2
loading y for sample 2
Sample(x=tensor([ 1.5761, -0.2646, -0.1489]), y=2)

Building a Dataloader

Now let’s see how we can use the Dataset class to build a DataLoader.

Metatensor’s DataLoader class is a wrapper around the PyTorch DataLoader class, and as such can be initialized with a Dataset object. It will also inherit all of the default arguments from the PyTorch DataLoader class.

in_memory_dataloader = DataLoader(in_memory_dataset)

We can now iterate over the DataLoader to access batches of samples from the dataset. With no arguments passed, the default batch size is 1 and the samples are not shuffled.

for batch in in_memory_dataloader:
    print(batch.y)

(0,)
(1,)
(2,)
(3,)
(4,)

As an alternative syntax, the data fields can be unpacked into separate variables in the for loop.

for x, y in in_memory_dataloader:
    print(x, y)

tensor([[-1.6455,  0.9555, -0.5801]]) (0,)
tensor([[-0.1646, -1.3758,  1.3614]]) (1,)
tensor([[ 1.5761, -0.2646, -0.1489]]) (2,)
tensor([[0.7162, 0.2067, 1.0703]]) (3,)
tensor([[-0.5449, -0.5098,  0.0966]]) (4,)

We can also pass arguments to the DataLoader constructor to change the batch size and shuffling of the samples.

in_memory_dataloader = DataLoader(in_memory_dataset, batch_size=2, shuffle=True)

for batch in in_memory_dataloader:
    print(batch.y)

(1, 4)
(0, 3)
(2,)

Data loaders for cross-validation

One can use the usual torch torch.utils.data.random_split() function to split a Dataset into train, validation, and test subsets for cross-validation purposes. DataLoader s can then be constructed for each subset.

# Perform a random train/val/test split of the Dataset,
# in the relative proportions (60% / 20% / 20%)
train_dataset, val_dataset, test_dataset = torch.utils.data.random_split(
    in_memory_dataset, [0.6, 0.2, 0.2]
)

# Construct DataLoaders for each subset
train_dataloader = DataLoader(train_dataset)
val_dataloader = DataLoader(val_dataset)
test_dataloader = DataLoader(test_dataset)

# As the Dataset was initialized with 5 samples, the split should be 3:1:1
print(f"Dataset size: {len(on_disk_dataset)}")
print(f"Training set size: {len(train_dataloader)}")
print(f"Validation set size: {len(val_dataloader)}")
print(f"Test set size: {len(test_dataloader)}")

Dataset size: 5
Training set size: 3
Validation set size: 1
Test set size: 1

Working with torch.Tensor and metatensor.torch.TensorMap

As the Dataset and DataLoader classes exist to interface metatensor and torch, let’s explore how they behave when using torch.Tensor and metatensor.torch.TensorMap objects as the data.

We’ll consider some dummy data consisting of the following fields:

• descriptor: a list of random TensorMap objects

• scalar: a list of random floats

• vector: a list of random torch Tensors

# Create a dummy descriptor as a TensorMap
descriptor = [
    TensorMap(
        keys=Labels(
            names=["key_1", "key_2"],
            values=torch.tensor([[1, 2]]),
        ),
        blocks=[
            TensorBlock(
                values=torch.randn((1, 3)),
                samples=Labels("sample_id", torch.tensor([[sample_id]])),
                components=[],
                properties=Labels("p", torch.tensor([[1], [4], [5]])),
            )
        ],
    )
    for sample_id in range(n_samples)
]

# Create dummy scalar and vectorial target properties as torch Tensors
scalar = [float(torch.rand(1, 1)) for _ in range(n_samples)]
vector = [torch.rand(1, 3) for _ in range(n_samples)]

# Build the Dataset
dataset = Dataset(
    scalar=scalar,
    vector=vector,
    descriptor=descriptor,
)
print(dataset[0])

Sample(scalar=0.19049400091171265, vector=tensor([[0.4452, 0.6630, 0.4448]]), descriptor=TensorMap with 1 blocks
keys: key_1  key_2
        1      2)

Merging samples in a batch

As is typically customary when working with torch Tensors, we want to vertically stack the samples in a minibatch into a single Tensor object. This allows passing a single Tensor object to a model, rather than a tuple of Tensor objects. In a similar way, sparse data stored in metatensor TensorMap objects can also be vertically stacked, i.e. joined along the samples axis, into a single TensorMap object.

The default collate_fn used by DataLoader (metatensor.learn.data.group_and_join()), vstacks (respectively joins along the samples axis) data fields that correspond torch.Tensor (respectively metatensor.torch.TensorMap). For all other data types, the data is left as tuple containing all samples in the current batch in order.

batch_size = 2
dataloader = DataLoader(dataset, batch_size=batch_size)

We can look at a single Batch object (i.e. a named tuple, returned by the DataLoader.__iter__()) to see this in action.

batch = next(iter(dataloader))

# TensorMaps for each sample in the batch joined along the samples axis
# into a single TensorMap
print("batch.descriptor =", batch.descriptor)

# `scalar` data are float objects, so are just grouped and returned in a tuple
print("batch.scalar =", batch.scalar)
assert len(batch.scalar) == batch_size

# `vector` data are torch Tensors, so are vertically stacked into a single
# Tensor
print("batch.vector =", batch.vector)

batch.descriptor = TensorMap with 1 blocks
keys: key_1  key_2
        1      2
batch.scalar = (0.19049400091171265, 0.6036732196807861)
batch.vector = tensor([[0.4452, 0.6630, 0.4448],
        [0.1524, 0.4974, 0.0150]])

Advanced functionality: IndexedDataset

What if we wanted to explicitly define the sample indexes used to store and access samples in the dataset? See the next tutorial, Using IndexedDataset, for more details!