| 1 | /**
|
| 2 | * @license
|
| 3 | * Copyright 2021 Google LLC. All Rights Reserved.
|
| 4 | * Licensed under the Apache License, Version 2.0 (the "License");
|
| 5 | * you may not use this file except in compliance with the License.
|
| 6 | * You may obtain a copy of the License at
|
| 7 | *
|
| 8 | * http://www.apache.org/licenses/LICENSE-2.0
|
| 9 | *
|
| 10 | * Unless required by applicable law or agreed to in writing, software
|
| 11 | * distributed under the License is distributed on an "AS IS" BASIS,
|
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 13 | * See the License for the specific language governing permissions and
|
| 14 | * limitations under the License.
|
| 15 | * =============================================================================
|
| 16 | */
|
| 17 | import { ENGINE } from '../engine';
|
| 18 | import { Einsum } from '../kernel_names';
|
| 19 | import { convertToTensor } from '../tensor_util_env';
|
| 20 | import { op } from './operation';
|
| 21 | /**
|
| 22 | * Tensor contraction over specified indices and outer product.
|
| 23 | *
|
| 24 | * `einsum` allows defining Tensors by defining their element-wise computation.
|
| 25 | * This computation is based on
|
| 26 | * [Einstein summation](https://en.wikipedia.org/wiki/Einstein_notation).
|
| 27 | *
|
| 28 | * Some special cases include:
|
| 29 | *
|
| 30 | * Matrix multiplication:
|
| 31 | * ```js
|
| 32 | * const x = tf.tensor2d([[1, 2, 3], [4, 5, 6]]);
|
| 33 | * const y = tf.tensor2d([[0, 1], [2, 3], [4, 5]]);
|
| 34 | * x.print();
|
| 35 | * y.print();
|
| 36 | * tf.einsum('ij,jk->ik', x, y).print();
|
| 37 | * ```
|
| 38 | *
|
| 39 | * Dot product:
|
| 40 | * ```js
|
| 41 | * const x = tf.tensor1d([1, 2, 3]);
|
| 42 | * const y = tf.tensor1d([0, 1, 2]);
|
| 43 | * x.print();
|
| 44 | * y.print();
|
| 45 | * tf.einsum('i,i->', x, y).print();
|
| 46 | * ```
|
| 47 | *
|
| 48 | * Batch dot product:
|
| 49 | * ```js
|
| 50 | * const x = tf.tensor2d([[1, 2, 3], [4, 5, 6]]);
|
| 51 | * const y = tf.tensor2d([[0, 1, 2], [3, 4, 5]]);
|
| 52 | * x.print();
|
| 53 | * y.print();
|
| 54 | * tf.einsum('bi,bi->b', x, y).print();
|
| 55 | * ```
|
| 56 | *
|
| 57 | * Outer prouduct:
|
| 58 | * ```js
|
| 59 | * const x = tf.tensor1d([1, 3, 5]);
|
| 60 | * const y = tf.tensor1d([2, 4, 6]);
|
| 61 | * x.print();
|
| 62 | * y.print();
|
| 63 | * tf.einsum('i,j->ij', x, y).print();
|
| 64 | * ```
|
| 65 | *
|
| 66 | * Matrix transpose:
|
| 67 | * ```js
|
| 68 | * const x = tf.tensor2d([[1, 2], [3, 4]]);
|
| 69 | * x.print();
|
| 70 | * tf.einsum('ij->ji', x).print();
|
| 71 | * ```
|
| 72 | *
|
| 73 | * Batch matrix transpose:
|
| 74 | * ```js
|
| 75 | * const x = tf.tensor3d([[[1, 2], [3, 4]], [[-1, -2], [-3, -4]]]);
|
| 76 | * x.print();
|
| 77 | * tf.einsum('bij->bji', x).print();
|
| 78 | * ```
|
| 79 | *
|
| 80 | * Limitations:
|
| 81 | *
|
| 82 | * This implementation of einsum has the following limitations:
|
| 83 | *
|
| 84 | * - Does not support >2 input tensors.
|
| 85 | * - Does not support duplicate axes for any given input tensor. E.g., equation
|
| 86 | * 'ii->' is not suppoted.
|
| 87 | * - The `...` notation is not supported.
|
| 88 | *
|
| 89 | * @param equation a string describing the contraction, in the same format as
|
| 90 | * [numpy.einsum](https://numpy.org/doc/stable/reference/generated/numpy.einsum.html).
|
| 91 | * @param tensors the input(s) to contract (each one a Tensor), whose shapes
|
| 92 | * should be consistent with equation.
|
| 93 | * @returns The output tensor.
|
| 94 | *
|
| 95 | * @doc {heading: 'Tensors', subheading: 'Matrices'}
|
| 96 | */
|
| 97 | export function einsum_(equation, ...tensors) {
|
| 98 | const $tensors = tensors.map((t, i) => convertToTensor(t, `tensors${i}`, 'einsum'));
|
| 99 | const attrs = { equation };
|
| 100 | return ENGINE.runKernel(Einsum, $tensors, attrs);
|
| 101 | }
|
| 102 | export const einsum = op({ einsum_ });
|
| 103 | //# sourceMappingURL=einsum.js.map |
| \ | No newline at end of file |