@tensorflow/tfjs-core/src/gradients.ts

Hardware-accelerated JavaScript library for machine intelligence

/**
 * @license
 * Copyright 2018 Google Inc. All Rights Reserved.
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 * http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 * =============================================================================
 */

import {CustomGradientFunc, ENGINE} from './engine';
import {Scalar, Tensor, Variable} from './tensor';
import {NamedTensorMap} from './tensor_types';
import {convertToTensor, convertToTensorArray} from './tensor_util_env';
import {TensorLike} from './types';
import * as util from './util';

/**
 * Provided `f(x)`, returns another function `g(x, dy?)`, which gives the
 * gradient of `f(x)` with respect to `x`.
 *
 * If `dy` is provided, the gradient of `f(x).mul(dy).sum()` with respect to
 * `x` is computed instead. `f(x)` must take a single tensor `x` and return a
 * single tensor `y`. If `f()` takes multiple inputs, use `tf.grads` instead.
 *
 * ```js
 * // f(x) = x ^ 2
 * const f = x => x.square();
 * // f'(x) = 2x
 * const g = tf.grad(f);
 *
 * const x = tf.tensor1d([2, 3]);
 * g(x).print();
 * ```
 *
 * ```js
 * // f(x) = x ^ 3
 * const f = x => x.pow(tf.scalar(3, 'int32'));
 * // f'(x) = 3x ^ 2
 * const g = tf.grad(f);
 * // f''(x) = 6x
 * const gg = tf.grad(g);
 *
 * const x = tf.tensor1d([2, 3]);
 * gg(x).print();
 * ```
 *
 * @param f The function f(x), to compute gradient for.
 */
/** @doc {heading: 'Training', subheading: 'Gradients'} */
function grad(f: (x: Tensor) => Tensor): (
    x: TensorLike|Tensor, dy?: TensorLike|Tensor) => Tensor {
  util.assert(
      util.isFunction(f), () => 'The f passed in grad(f) must be a function');
  return (x: TensorLike|Tensor, dy?: TensorLike|Tensor): Tensor => {
    // x can be of any dtype, thus null as the last argument.
    const $x = convertToTensor(x, 'x', 'tf.grad', null);
    const $dy: Tensor =
        (dy != null) ? convertToTensor(dy, 'dy', 'tf.grad') : null;
    return ENGINE.tidy(() => {
      const {value, grads} = ENGINE.gradients(() => f($x), [$x], $dy);
      if ($dy != null) {
        util.assertShapesMatch(
            value.shape, $dy.shape,
            'The shape of dy passed in grad(f)(x, dy) must match the shape ' +
                'returned by f(x)');
      }
      checkGrads(grads);
      return grads[0];
    });
  };
}

/**
 * Provided `f(x1, x2,...)`, returns another function `g([x1, x2,...], dy?)`,
 * which gives an array of gradients of `f()` with respect to each input
 * [`x1`,`x2`,...].
 *
 * If `dy` is passed when calling `g()`, the gradient of
 * `f(x1,...).mul(dy).sum()` with respect to each input is computed instead.
 * The provided `f` must take one or more tensors and return a single tensor
 * `y`. If `f()` takes a single input, we recommend using `tf.grad` instead.
 *
 * ```js
 * // f(a, b) = a * b
 * const f = (a, b) => a.mul(b);
 * // df / da = b, df / db = a
 * const g = tf.grads(f);
 *
 * const a = tf.tensor1d([2, 3]);
 * const b = tf.tensor1d([-2, -3]);
 * const [da, db] = g([a, b]);
 * console.log('da');
 * da.print();
 * console.log('db');
 * db.print();
 * ```
 *
 * @param f The function `f(x1, x2,...)` to compute gradients for.
 */
/** @doc {heading: 'Training', subheading: 'Gradients'} */
function grads(f: (...args: Tensor[]) => Tensor): (
    args: Array<Tensor|TensorLike>, dy?: Tensor|TensorLike) => Tensor[] {
  util.assert(
      util.isFunction(f), () => 'The f passed in grads(f) must be a function');
  return (args: Array<Tensor|TensorLike>, dy?: Tensor|TensorLike): Tensor[] => {
    util.assert(
        Array.isArray(args),
        () => 'The args passed in grads(f)(args) must be an array ' +
            'of `Tensor`s or `TensorLike`s');
    // args can be of any dtype, thus null as the last argument.
    const $args = convertToTensorArray(args, 'args', 'tf.grads', null);
    const $dy: Tensor =
        (dy != null) ? convertToTensor(dy, 'dy', 'tf.grads') : null;
    return ENGINE.tidy(() => {
      const {value, grads} = ENGINE.gradients(() => f(...$args), $args, $dy);
      if ($dy != null) {
        util.assertShapesMatch(
            value.shape, $dy.shape,
            'The shape of dy passed in grads(f)([x1,...], dy) must ' +
                'match the shape returned by f([x1,...])');
      }
      checkGrads(grads);
      return grads;
    });
  };
}

/**
 * Like `tf.grad`, but also returns the value of `f()`. Useful when `f()`
 * returns a metric you want to show.
 *
 * The result is a rich object with the following properties:
 * - grad: The gradient of `f(x)` w.r.t `x` (result of `tf.grad`).
 * - value: The value returned by `f(x)`.
 *
 * ```js
 * // f(x) = x ^ 2
 * const f = x => x.square();
 * // f'(x) = 2x
 * const g = tf.valueAndGrad(f);
 *
 * const x = tf.tensor1d([2, 3]);
 * const {value, grad} = g(x);
 *
 * console.log('value');
 * value.print();
 * console.log('grad');
 * grad.print();
 * ```
 */
/** @doc {heading: 'Training', subheading: 'Gradients'} */
function valueAndGrad<I extends Tensor, O extends Tensor>(f: (x: I) => O): (
    x: I, dy?: O) => {
  value: O;
  grad: I;
} {
  util.assert(
      util.isFunction(f),
      () => 'The f passed in valueAndGrad(f) must be a function');
  return (x: I, dy?: O) => {
    util.assert(
        x instanceof Tensor,
        () => 'The x passed in valueAndGrad(f)(x) must be a tensor');
    util.assert(
        dy == null || dy instanceof Tensor,
        () => 'The dy passed in valueAndGrad(f)(x, dy) must be a tensor');
    const {grads, value} = ENGINE.gradients(() => f(x), [x], dy);
    checkGrads(grads);
    return {grad: grads[0] as I, value: value as O};
  };
}

/**
 * Like `tf.grads`, but returns also the value of `f()`. Useful when `f()`
 * returns a metric you want to show.
 *
 * The result is a rich object with the following properties:
 * - grads: The gradients of `f()` w.r.t each input (result of `tf.grads`).
 * - value: The value returned by `f(x)`.
 *
 * ```js
 * // f(a, b) = a * b
 * const f = (a, b) => a.mul(b);
 * // df/da = b, df/db = a
 * const g = tf.valueAndGrads(f);
 *
 * const a = tf.tensor1d([2, 3]);
 * const b = tf.tensor1d([-2, -3]);
 * const {value, grads} = g([a, b]);
 *
 * const [da, db] = grads;
 *
 * console.log('value');
 * value.print();
 *
 * console.log('da');
 * da.print();
 * console.log('db');
 * db.print();
 * ```
 */
/** @doc {heading: 'Training', subheading: 'Gradients'} */
function valueAndGrads<O extends Tensor>(f: (...args: Tensor[]) => O): (
    args: Tensor[], dy?: O) => {
  grads: Tensor[];
  value: O;
} {
  util.assert(
      util.isFunction(f),
      () => 'The f passed in valueAndGrads(f) must be a function');
  return (args: Tensor[], dy?: O) => {
    util.assert(
        Array.isArray(args) && args.every(arg => arg instanceof Tensor),
        () => 'The args passed in valueAndGrads(f)(args) must be array of ' +
            'tensors');
    util.assert(
        dy == null || dy instanceof Tensor,
        () => 'The dy passed in valueAndGrads(f)(args, dy) must be a tensor');
    const res = ENGINE.gradients(() => f(...args), args, dy);
    if (dy != null) {
      util.assertShapesMatch(
          res.value.shape, dy.shape,
          'The shape of dy passed in valueAndGrads(f)([x1,...], dy) must ' +
              'match the shape returned by f([x1,...])');
    }
    checkGrads(res.grads);
    return res;
  };
}

/**
 * Computes and returns the gradient of f(x) with respect to the list of
 * trainable variables provided by `varList`. If no list is provided, it
 * defaults to all trainable variables.
 *
 * ```js
 * const a = tf.variable(tf.tensor1d([3, 4]));
 * const b = tf.variable(tf.tensor1d([5, 6]));
 * const x = tf.tensor1d([1, 2]);
 *
 * // f(a, b) = a * x ^ 2 + b * x
 * const f = () => a.mul(x.square()).add(b.mul(x)).sum();
 * // df/da = x ^ 2, df/db = x
 * const {value, grads} = tf.variableGrads(f);
 *
 * Object.keys(grads).forEach(varName => grads[varName].print());
 * ```
 *
 * @param f The function to execute. f() should return a scalar.
 * @param varList The list of variables to compute the gradients with respect
 *     to. Defaults to all trainable variables.
 * @returns An object with the following keys and values:
 *   - `value`: The value of the function `f`.
 *   - `grads`: A map from the names of the variables to the gradients.
 *     If the `varList` argument is provided explicitly and contains a subset of
 *     non-trainable variables, this map in the return value will contain keys
 *     that map the names of the non-trainable variables to `null`.
 */
/** @doc {heading: 'Training', subheading: 'Gradients'} */
function variableGrads(f: () => Scalar, varList?: Variable[]):
    {value: Scalar, grads: NamedTensorMap} {
  util.assert(
      util.isFunction(f),
      () => 'The f passed in variableGrads(f) must be a function');
  util.assert(
      varList == null ||
          Array.isArray(varList) && varList.every(v => v instanceof Variable),
      () =>
          'The varList passed in variableGrads(f, varList) must be an array ' +
          'of variables');

  const specifiedVarList = varList != null;
  if (!specifiedVarList) {
    // Get all of the trainable variables.
    varList = [];
    for (const varName in ENGINE.registeredVariables) {
      varList.push(ENGINE.registeredVariables[varName]);
    }
  }

  const specifiedNonTrainable: Variable[] =
      specifiedVarList ? varList.filter(variable => !variable.trainable) : null;

  // Prune non-trainable variables.
  const originalVarCount = varList.length;
  varList = varList.filter(variable => variable.trainable);
  util.assert(
      varList.length > 0,
      () => `variableGrads() expects at least one of the input variables to ` +
          `be trainable, but none of the ${originalVarCount} variables is ` +
          `trainable.`);

  const allowNoGradients = true;
  const {value, grads} = ENGINE.gradients(f, varList, null, allowNoGradients);

  util.assert(
      grads.some(g => g != null),
      () => 'Cannot find a connection between any variable and the result of ' +
          'the loss function y=f(x). Please make sure the operations that ' +
          'use variables are inside the function f passed to minimize().');
  util.assert(
      value.rank === 0,
      () => `The f passed in variableGrads(f) must return a scalar, but it ` +
          `returned a rank-${value.rank} tensor`);

  const namedGrads: NamedTensorMap = {};
  varList.forEach((v, i) => {
    if (grads[i] != null) {
      namedGrads[v.name] = grads[i];
    }
  });
  if (specifiedNonTrainable != null) {
    // If varList is explicitly provided and contains non-trainable values,
    // add them to the returned gradients with `null` values.
    specifiedNonTrainable.forEach(v => namedGrads[v.name] = null);
  }
  return {value, grads: namedGrads};
}

/**
 * Overrides the gradient computation of a function `f`.
 *
 * Takes a function
 * `f(...inputs, save) => {value: Tensor, gradFunc: (dy, saved) => Tensor[]}`
 * and returns another function `g(...inputs)` which takes the same inputs as
 * `f`. When called, `g` returns `f().value`. In backward mode, custom gradients
 * with respect to each input of `f` are computed using `f().gradFunc`.
 *
 * The `save` function passsed to `f` should be used for saving tensors needed
 * in the gradient. And the `saved` passed to the `gradFunc` is a
 * `NamedTensorMap`, which contains those saved tensor.
 *
 * ```js
 * const customOp = tf.customGrad((x, save) => {
 *   // Save x to make sure it's available later for the gradient.
 *   save([x]);
 *   // Override gradient of our custom x ^ 2 op to be dy * abs(x);
 *   return {
 *     value: x.square(),
 *     // Note `saved.x` which points to the `x` we saved earlier.
 *     gradFunc: (dy, saved) => [dy.mul(saved[0].abs())]
 *   };
 * });
 *
 * const x = tf.tensor1d([-1, -2, 3]);
 * const dx = tf.grad(x => customOp(x));
 *
 * console.log(`f(x):`);
 * customOp(x).print();
 * console.log(`f'(x):`);
 * dx(x).print();
 * ```
 *
 * @param f The function to evaluate in forward mode, which should return
 *     `{value: Tensor, gradFunc: (dy, saved) => Tensor[]}`, where `gradFunc`
 *     returns the custom gradients of `f` with respect to its inputs.
 */
/** @doc {heading: 'Training', subheading: 'Gradients'} */
function customGrad<T extends Tensor>(f: CustomGradientFunc<T>):
    (...args: Tensor[]) => T {
  return ENGINE.customGrad(f);
}

function checkGrads(grads: Tensor[]) {
  const numNullGradients = grads.filter(g => g == null).length;
  if (numNullGradients > 0) {
    throw new Error(
        `Cannot compute gradient of y=f(x) with respect to x. Make sure that
    the f you passed encloses all operations that lead from x to y.`);
  }
}

export {
  customGrad,
  variableGrads,
  valueAndGrad,
  valueAndGrads,
  grad,
  grads,
};