Using IndexedDataset

import os

import torch

from metatensor.learn.data import DataLoader, Dataset, IndexedDataset

Review of the standard Dataset

The previous tutorial, Datasets and data loaders, showed how to define a Dataset able to handle both torch tensor and metatensor TensorMap. We saw that in-memory, on-disk, or mixed in-memory/on-disk datasets can be defined. DataLoaders are then defined on top of these Dataset objects.

In all cases, however, each data sample is accessed by a numeric integer index, which ranges from 0 to len(dataset) - 1. Let’s use a simple example to review this.

Again let’s define some dummy data as before. Our x data as a list of random tensors, and our y data as a list of integers that enumerate the samples.

For the purposes of this tutorial, we will only focus on an in-memory dataset, though the same principles apply to on-disk and mixed datasets.

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

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

A sample is accessed by its numeric index. As the length of the lists passed as kwargs is 5, both for x and y, the valid indices are [0, 1, 2, 3, 4].

Let’s retrieve the 4th sample (index 3) and print it. The value of the “y” data field should be 3.

print(dataset[3])

Sample(x=tensor([-1.2287, -0.6355, -0.6601]), y=3)

What if we wanted to access samples by something other than an integer index part of a continuous range?

For instance, what if we wanted to access samples by:

1. a string id, or other arbitrary hashable object?

2. an integer index that is not defined inside a continuous range?

In these cases, we can use an IndexedDataset instead.

IndexedDataset

First let’s define a Dataset where the samples are indexed by arbitrary unique indexes, such as strings, integers, and tuples.

Suppose the unique indexes for our 5 samples are:

sample_id = [
    "cat",
    4,
    ("small", "cow"),
    "dog",
    0,
]

# Build an IndexedDataset, specifying the unique sample indexes with ``sample_id``
dataset = IndexedDataset(
    x=x_data,
    y=y_data,
    sample_id=sample_id,
)

Now, when we access the dataset, we can access samples by their unique sample index using the get_sample method. This method takes a single argument, the sample index, and returns the corresponding sample.

print(dataset.get_sample("dog"))
print(dataset.get_sample(4))
print(dataset.get_sample(("small", "cow")))

Sample(sample_id='dog', x=tensor([-1.2287, -0.6355, -0.6601]), y=3)
Sample(sample_id=4, x=tensor([-1.2476, -1.0920,  1.4115]), y=1)
Sample(sample_id=('small', 'cow'), x=tensor([ 1.7618, -0.5364,  0.3124]), y=2)

Note that using __getitem__, i.e. dataset[4], will return the sample passed to the constructor at position 5. In this case, the sample indexes map to the numeric indices as follows:

0. "cat"

1. 4

2. ("small", "cow")

3. "dog"

4. 0

Thus, accessing the unique sample index "cat" can be done equivalently with either of:

print(dataset[0])
print(dataset.get_sample("cat"))

Sample(sample_id='cat', x=tensor([-1.1320, -1.3363, -0.0708]), y=0)
Sample(sample_id='cat', x=tensor([-1.1320, -1.3363, -0.0708]), y=0)

Note that the named tuple returned in both cases contains the unique sample index as the sample_id field, which precedes all other data fields. This is in contrast to the standard Dataset, which only returns the passed data fields and not the index.

A DataLoader can be constructed on top of an IndexedDataset in the same way as a Dataset. Batches are accessed by iterating over the DataLoader, though this time the Batch named tuple returned by the data loader will contain the unique sample indexes sample_id as the first field.

dataloader = DataLoader(dataset, batch_size=2)

# Iterate over batches
for batch in dataloader:
    print(batch)

Batch(sample_id=('cat', 4), x=tensor([[-1.1320, -1.3363, -0.0708],
        [-1.2476, -1.0920,  1.4115]]), y=(0, 1))
Batch(sample_id=(('small', 'cow'), 'dog'), x=tensor([[ 1.7618, -0.5364,  0.3124],
        [-1.2287, -0.6355, -0.6601]]), y=(2, 3))
Batch(sample_id=(0,), x=tensor([[-0.3961,  1.2469,  0.3087]]), y=(4,))

As before, we can create separate variables in the iteration pattern

for ids, x, y in dataloader:
    print(ids, x, y)

('cat', 4) tensor([[-1.1320, -1.3363, -0.0708],
        [-1.2476, -1.0920,  1.4115]]) (0, 1)
(('small', 'cow'), 'dog') tensor([[ 1.7618, -0.5364,  0.3124],
        [-1.2287, -0.6355, -0.6601]]) (2, 3)
(0,) tensor([[-0.3961,  1.2469,  0.3087]]) (4,)

On-disk IndexedDataset with arbitrary sample indexes

When defining an IndexedDataset with data fields on-disk, i.e. to be loaded lazily, the sample indexes passed as the sample_id kwarg to the constructor are used as the arguments to the load function.

To demonstrate this, as we did in the previous tutorial, let’s save the x data to disk and build a mixed in-memory/on-disk IndexedDataset.

For instance, the below code will save sone x data for the sample "dog" at relative path "data/x_dog.pt".

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

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

We can now define a load function to load data from disk. This should take the unique sample index as a single argument, and return the corresponding data in memory.

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")

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

mixed_dataset = IndexedDataset(x=load_x, y=y_data, sample_id=sample_id)
print(mixed_dataset.get_sample("dog"))
print(mixed_dataset.get_sample(("small", "cow")))

loading x for sample dog
/home/runner/work/metatensor/metatensor/python/examples/learn/2-indexed-dataset.py:174: 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")
Sample(sample_id='dog', x=tensor([-1.2287, -0.6355, -0.6601]), y=3)
loading x for sample ('small', 'cow')
Sample(sample_id=('small', 'cow'), x=tensor([ 1.7618, -0.5364,  0.3124]), y=2)

Using an IndexedDataset: subset integer ranges

One could also define an IndexedDataset where the samples indices are integers forming a possibly shuffled and non-continuous subset of a larger continuous range of numeric indices.

For instance, imagine we have a global Dataset of 1000 samples, with indices [0, …, 999], but only want to build a dataset for samples with indices [4, 7, 200, 5, 999], in that order. We can pass these indices kwarg sample_id.

# Build an IndexedDataset, specifying the subset sample indexes in a specific order
sample_id = [4, 7, 200, 5, 999]
dataset = IndexedDataset(x=x_data, y=y_data, sample_id=sample_id)

Now, when we access the dataset, we can access samples by their unique sample index using the get_sample method. This method takes a single argument, the sample index, and returns the corresponding sample.

Again, the numeric index can be used equivalently to access the sample, and again note that the Sample named tuple includes the sample_id field.

# These return the same sample
print(dataset.get_sample(5))
print(dataset[4])

Sample(sample_id=5, x=tensor([-1.2287, -0.6355, -0.6601]), y=3)
Sample(sample_id=999, x=tensor([-0.3961,  1.2469,  0.3087]), y=4)

And finally, the DataLoader behaves as expected:

dataloader = DataLoader(dataset, batch_size=2)

for batch in dataloader:
    print(batch)

Batch(sample_id=(4, 7), x=tensor([[-1.1320, -1.3363, -0.0708],
        [-1.2476, -1.0920,  1.4115]]), y=(0, 1))
Batch(sample_id=(200, 5), x=tensor([[ 1.7618, -0.5364,  0.3124],
        [-1.2287, -0.6355, -0.6601]]), y=(2, 3))
Batch(sample_id=(999,), x=tensor([[-0.3961,  1.2469,  0.3087]]), y=(4,))