First steps with metatensor

This tutorial explores how data is stored inside metatensor’s TensorMap, and how to access the associated metadata. This is a companion to the core classes overview page of this documentation, presenting the same concepts with code examples.

To this end, we will need some data in metatensor format, which for the sake of simplicity will be loaded from a file. The code used to generate this file can be found below:

Show the code used to generate the spherical-expansion.npz file, or use the link to download it

The data was generated with rascaline, a package to compute atomistic representations for machine learning applications.

import ase
from rascaline import SphericalExpansion

import metatensor

co2 = ase.Atoms(
    positions=[(0, 0, 0), (-0.2, -0.65,  0.94), (0.2, 0.65, -0.94)],

calculator = SphericalExpansion(
    radial_basis={"Gto": {}},
    cutoff_function={"ShiftedCosine": {"width": 0.5}},

descriptor = calculator.compute(co2, gradients=["positions"])"spherical-expansion.npz", descriptor)

The TensorMap stored in the file contains a machine learning representation (the spherical expansion) of all the atoms in a CO2 molecule. You don’t need to know anything the spherical expansion to follow this tutorial!

import ase
import ase.visualize.plot
import matplotlib.pyplot as plt

import metatensor

For reference, we are working with a representation of this CO2 molecule:

co2 = ase.Atoms(
    positions=[(0, 0, 0), (-0.2, -0.65, 0.94), (0.2, 0.65, -0.94)],

fig, ax = plt.subplots(figsize=(3, 3))
ase.visualize.plot.plot_atoms(co2, ax)
1 first steps

The main entry point: TensorMap

We’ll start by loading our data with metatensor.load(). The tensor returned by this function is a TensorMap, the core class of metatensor.

tensor = metatensor.load("spherical-expansion.npz")
<class 'metatensor.tensor.TensorMap'>

Looking at the tensor tells us that it is composed of 12 blocks, each associated with a key:

TensorMap with 12 blocks
keys: o3_lambda  o3_sigma  center_type  neighbor_type
          0         1           6             6
          1         1           6             6
          2         1           6             6
          0         1           6             8
          1         1           6             8
          2         1           6             8
          0         1           8             6
          1         1           8             6
          2         1           8             6
          0         1           8             8
          1         1           8             8
          2         1           8             8

We can see that here, the keys of the TensorMap have four named dimensions. Two of these are used to describe the behavior of the data under spatial transformations (rotations and inversions in the O3 group):

  • o3_lambda, indicating the character of o3 irreducible representation this block is following. In general, a block with o3_lambda=3 will transform under rotations like a l=3 spherical harmonics.

  • o3_sigma, which describe the behavior of the data under inversion symmetry. Here all blocks have o3_sigma=1, meaning we only have data with the usual inversion symmetry (o3_sigma=-1 would be used for pseudo-tensors);

And the other two are related to the composition of the system:

  • center_type represents the atomic type of the central atom in consideration. For CO2, we have both carbons (type 6) and oxygens (type 8);

  • neighbor_type represents the atomic type of the neighbor atoms considered by the machine learning representation, in this case it takes the values 6 and 8 as well.

These keys can be accessed with TensorMap.keys, and they are an instance of the Labels class:

keys = tensor.keys
<class 'metatensor.labels.Labels'>

Labels to store metadata

One of the main goals of metatensor is to be able to store both data and metadata together. We’ve just encountered the first example of this metadata as the TensorMap keys! In general, most metadata will be stored in the Labels class. Let’s explore this class a bit.

As already mentioned, Labels can have multiple dimensions, and each dimension has a name. We can look at all the dimension names simultaneously with Labels.names():

['o3_lambda', 'o3_sigma', 'center_type', 'neighbor_type']

Labels then contains multiple entries, each entry being described by a set of integer values, one for each dimension of the labels.

[[0 1 6 6]
 [1 1 6 6]
 [2 1 6 6]
 [0 1 6 8]
 [1 1 6 8]
 [2 1 6 8]
 [0 1 8 6]
 [1 1 8 6]
 [2 1 8 6]
 [0 1 8 8]
 [1 1 8 8]
 [2 1 8 8]]

We can access all the values taken by a given dimension/column in the labels with Labels.column() or by indexing with a string:

[0 1 2 0 1 2 0 1 2 0 1 2]
[6 6 6 6 6 6 8 8 8 8 8 8]

We can also access individual entries in the labels by iterating over them or indexing with an integer:

print("Entries with o3_lambda=2:")
for entry in keys:
    if entry["o3_lambda"] == 2:
        print("    ", entry)

print("\nEntry at index 3:")
print("    ", keys[3])
Entries with o3_lambda=2:
     LabelsEntry(o3_lambda=2, o3_sigma=1, center_type=6, neighbor_type=6)
     LabelsEntry(o3_lambda=2, o3_sigma=1, center_type=6, neighbor_type=8)
     LabelsEntry(o3_lambda=2, o3_sigma=1, center_type=8, neighbor_type=6)
     LabelsEntry(o3_lambda=2, o3_sigma=1, center_type=8, neighbor_type=8)

Entry at index 3:
     LabelsEntry(o3_lambda=0, o3_sigma=1, center_type=6, neighbor_type=8)

TensorBlock to store the data

Each entry in the TensorMap.keys is associated with a TensorBlock, which contains the actual data and some additional metadata. We can extract the block from a key by indexing our TensorMap, or with the TensorMap.block()

# this is equivalent to `block = tensor[tensor.keys[0]]`
block = tensor[0]

block = tensor.block(o3_lambda=1, center_type=8, neighbor_type=6)

    samples (2): ['system', 'atom']
    components (3): ['o3_mu']
    properties (5): ['n']
    gradients: ['positions']

Each block contains some data, stored inside the TensorBlock.values. Here, the values contains the different coefficients of the spherical expansion, i.e. our atomistic machine learning representation.

The problem with this array is that we do not know what the different numbers correspond to: different libraries might be using different convention and storage order, and one has to read documentation carefully if they want to use this kind of data. Metatensor helps by making this data self-describing; by attaching metadata to each element of the array indicating what exactly we are working with.

[[[ 2.41688320e-02  1.37159979e-01  4.01218353e-02 -1.59115730e-04
  [-3.49518493e-02 -1.98354431e-01 -5.80223464e-02  2.30105825e-04
  [ 7.43656369e-03  4.22030705e-02  1.23451801e-02 -4.89586862e-05

 [[-2.41688320e-02 -1.37159979e-01 -4.01218353e-02  1.59115730e-04
  [ 3.49518493e-02  1.98354431e-01  5.80223464e-02 -2.30105825e-04
  [-7.43656369e-03 -4.22030705e-02 -1.23451801e-02  4.89586862e-05

The metadata is attached to the different array axes, and stored in Labels. The array must have at least two axes but can have more if required. Here, we have three:

(2, 3, 5)

The first dimension of the values array is described by the TensorBlock.samples labels, and correspond to what is being described. This follows the usual convention in machine learning, using the different rows of the array to store separate samples/observations.

Here, since we are working with a per-atom representation, the samples contain the index of the structure and atomic center in this structure. Since we are looking at a block for center_type=8, we have two samples, one for each oxygen atom in our single CO2 molecule.

    system  atom
      0      1
      0      2

The last dimension of the values array is described by the labels, and correspond to how we are describing our subject. Here, we are using a radial basis, indexed by an integer n:


Finally, each intermediate dimension of the values array is described by one set of TensorBlock.components labels. These dimensions correspond to one or more vectorial components in the data. Here the only component corresponds to the different \(m\) number in spherical harmonics \(Y_l^m\), going from -1 to 1 since we are looking at the block for o3_lambda = 1:


All this metadata allow us to know exactly what each entry in the values corresponds to. For example, we can see that the value at position (1, 0, 3) corresponds to:

  • the center at index 2 inside the structure at index 0;

  • the m=-1 part of the spherical harmonics;

  • the coefficients on the n=3 radial basis function.

print("value =", block.values[1, 0, 3])
print("sample =", block.samples[1])
print("component =", block.components[0][0])
print("property =",[3])
value = 0.00015911573016680835
sample = LabelsEntry(system=0, atom=2)
component = LabelsEntry(o3_mu=-1)
property = LabelsEntry(n=3)

Wrapping it up


Illustration of the structure of a TensorMap, with multiple keys and blocks.

To summarize this tutorial, we saw that a TensorMap contains multiple TensorBlock, each associated with a key. The key describes the block, and what kind of data will be found inside.

The blocks contains the actual data, and multiple set of metadata, one for each axis of the data array.

  • The rows are described by samples labels, which describe what is being stored;

  • the (generalized) columns are described by properties, which describe how the data is being represented;

  • Additional axes of the array correspond to vectorial components in the data.

All the metadata is stored inside Labels, where each entry is described by the integer values is takes along some named dimensions.

For a more visual approach to this data organization, you can also read the core classes overview.

We have learned how metatensor organizes its data, and what makes it a “self describing data format”. In the next tutorial, we will explore what makes metatensor TensorMap a “sparse data format”.

Total running time of the script: (0 minutes 1.206 seconds)

Gallery generated by Sphinx-Gallery