Einshape: DSL-based reshaping library for JAX and other frameworks.

The jnp.einsum op provides a DSL-based unified interface to matmul and tensordot ops. This einshape library is designed to offer a similar DSL-based approach to unifying reshape, squeeze, expand_dims, and transpose operations.

Some examples:

• einshape("n->n111", x) is equivalent to expand_dims(x, axis=1) three times
• einshape("a1b11->ab", x) is equivalent to squeeze(x, axis=[1,3,4])
• einshape("nhwc->nchw", x) is equivalent to transpose(x, perm=[0,3,1,2])
• einshape("mnhwc->(mn)hwc", x) is equivalent to a reshape combining the two leading dimensions
• einshape("(mn)hwc->mnhwc", x, n=batch_size) is equivalent to a reshape splitting the leading dimension into two, using kwargs (m or n or both) to supply the necessary additional shape information
• einshape("mn...->(mn)...", x) combines the two leading dimensions without knowing the rank of x
• einshape("n...->n(...)", x) performs a 'batch flatten'
• einshape("ij->ijk", x, k=3) inserts a trailing dimension and tiles along it
• einshape("ij->i(nj)", x, n=3) tiles along the second dimension

See jax_ops.py for the JAX implementation of the einshape function. Alternatively, the parser and engine are exposed in engine.py allowing analogous implementations in TensorFlow or other frameworks.

Installation

Einshape can be installed with the following command:

pip3 install git+https://github.com/deepmind/einshape

Einshape will work with either Jax or TensorFlow. To allow for that it does not list either as a requirement, so it is necessary to ensure that Jax or TensorFlow is installed separately.

Usage

Jax version:

from einshape import jax_einshape as einshape
from jax import numpy as jnp

a = jnp.array([[1, 2], [3, 4]])
b = einshape("ij->(ij)", a)
# b is [1, 2, 3, 4]

TensorFlow version:

from einshape import tf_einshape as einshape
import tensorflow as tf

a = tf.constant([[1, 2], [3, 4]])
b = einshape("ij->(ij)", a)
# b is [1, 2, 3, 4]

Numpy version:

from einshape import numpy_einshape as einshape
import numpy as np

a = np.array([[1, 2], [3, 4]])
b = einshape("ij->(ij)", a)
# b is [1, 2, 3, 4]

Understanding einshape equations

An einshape equation is always of the form {lhs}->{rhs}, where {lhs} and {rhs} both stand for expressions. An expression represents the axes of an array; the relationship between two expressions illustrate how an array should be transformed.

An expression is a non-empty sequence of the following elements:

Index name

A single letter a-z, representing one axis of an array.

For example, the expressions ab and jq both represent an array of rank 2.

Every index name that is present on the left-hand side of an equation must also be present on the right-hand side. So, ab->a is not a valid equation, but a->ba is valid (and will tile a vector b times).

Ellipsis

..., representing any axes of an array that are not otherwise represented in the expression. This is similar to the use of -1 as an axis in a reshape operation.

For example, a...b can represent any array of rank 2 or more: a will refer to the first axis and b to the last. The equation ...ab->...ba will swap the last two axes of an array.

An expression may not include more than one ellipsis (because that would be ambiguous). Like an index name, an ellipsis must be present in both halves of an equation or neither.

Group

({components}), where components is a sequence of index names and ellipsis elements. The entire group corresponds to a single axis of the array; the group's components represent factors of the axis size. This can be used to reshape an axis into many axes. All the factors except at most one must be specified using keyword arguments.

For example, einshape('(ab)->ab', x, a=10) reshapes an array of rank 1 (whose length must be a multiple of 10) into an array of rank 2 (whose first dimension is of length 10).

Groups may not be nested.

Unit

The digit 1, representing a single axis of length 1. This is useful for expanding and squeezing unit dimensions.

For example, the equation 1...->... squeezes a leading axis (which must have length one).

Disclaimer

This is not an official Google product.