我对换能器的理解是否正确？

Question

让我们从定义开始:A transducer是一个函数,它接受一个reducer函数并返回一个reducer函数.

A reducer是一个二元函数,它接受累加器和值并返回累加器.一个reducer可以用一个reduce函数执行(注意:所有函数都是curry但是我已经把它和它的定义pipe以及compose为了便于阅读 - 你可以在现场演示中看到它们):

const reduce = (reducer, init, data) => {
  let result = init;
  for (const item of data) {
    result = reducer(result, item);
  }
  return result;
}

随着reduce我们可以实现map和filter功能:

const mapReducer = xf => (acc, item) => [...acc, xf(item)];
const map = (xf, arr) => reduce(mapReducer(xf), [], arr);

const filterReducer = predicate => (acc, item) => predicate(item) ?
  [...acc, item] :
  acc;
const filter = (predicate, arr) => reduce(filterReducer(predicate), [], arr);

我们可以看到,这些函数之间存在一些相似之处map,filter并且这两个函数仅适用于数组.另一个缺点是,当我们编写这两个函数时,在每个步骤中都会创建一个临时数组,并将其传递给另一个函数.

const even = n => n % 2 === 0;
const double = n => n * 2;

const doubleEven = pipe(filter(even), map(double));

doubleEven([1,2,3,4,5]);
// first we get [2, 4] from filter
// then final result: [4, 8]

传感器帮助我们解决了这些问题:当我们使用传感器时,没有创建临时阵列,我们可以将我们的函数概括为不仅适用于数组.传感器需要一个transduce功能才能工作传感器通常通过传递给transduce函数来执行:

const transduce = (xform, iterator, init, data) =>
  reduce(xform(iterator), init, data);

const mapping = (xf, reducer) => (acc, item) => reducer(acc, xf(item));

const filtering = (predicate, reducer) => (acc, item) => predicate(item) ?
  reducer(acc, item) :
  acc;

const arrReducer = (acc, item) => [...acc, item];

const transformer = compose(filtering(even), mapping(double));

const performantDoubleEven = transduce(transformer, arrReducer, [])

performantDoubleEven([1, 2, 3, 4, 5]); // -> [4, 8] with no temporary arrays created

我们甚至可以定义数组map和filter使用transducer因为它是如此可组合:

const map = (xf, data) => transduce(mapping(xf), arrReducer, [], data);

const filter = (predicate, data) => transduce(filtering(predicate), arrReducer, [], data);

现场版如果你想运行代码 - > https://runkit.com/marzelin/transducers

我的推理有意义吗？

Answer 1

你的理解是正确的，但不完整。

除了您描述的概念之外，传感器还可以执行以下操作：

• 支持提前退出语义
• 支持完成语义
• 有状态
• 支持阶跃函数的初始值。

例如，JavaScript 中的实现需要执行以下操作：

// Ensure reduce preserves early termination
let called = 0;
let updatesCalled = map(a => { called += 1; return a; });
let hasTwo = reduce(compose(take(2), updatesCalled)(append), [1,2,3]).toString();
console.assert(hasTwo === '1,2', hasTwo);
console.assert(called === 2, called);

这里因为调用了take归约操作而提前退出。

它需要能够（可选）调用不带初始值参数的步骤函数：

// handles lack of initial value
let mapDouble = map(n => n * 2);
console.assert(reduce(mapDouble(sum), [1,2]) === 6);

这里，不带参数的调用sum返回加性恒等式（零）以作为归约的种子。

为了实现这一点，这里有一个辅助函数：

const addArities = (defaultValue, reducer) => (...args) => {
  switch (args.length) {
    case 0: return typeof defaultValue === 'function' ? defaultValue() : defaultValue;
    case 1: return args[0];
    default: return reducer(...args);
  }
};

这需要一个初始值（或可以提供初始值的函数）和一个减速器作为种子：

const sum = addArities(0, (a, b) => a + b);

现在sum具有正确的语义，这也是append第一个示例中的定义方式。对于有状态转换器，请查看take（包括辅助函数）：

// Denotes early completion
class _Wrapped {
  constructor (val) { this[DONE] = val }
};

const isReduced = a => a instanceof _Wrapped;
// ensures reduced for bubbling
const reduced = a => a instanceof _Wrapped ? a : new _Wrapped(a);
const unWrap = a => isReduced(a) ? a[DONE] : a;

const enforceArgumentContract = f => (xform, reducer, accum, input, state) => {
  // initialization
  if (!exists(input)) return reducer();
  // Early termination, bubble
  if (isReduced(accum)) return accum;
  return f(xform, reducer, accum, input, state);
};

/*
 * factory
 *
 * Helper for creating transducers.
 *
 * Takes a step process, intial state and returns a function that takes a
 * transforming function which returns a transducer takes a reducing function,
 * optional collection, optional initial value. If collection is not passed
 * returns a modified reducing function, otherwise reduces the collection.
 */
const factory = (process, initState) => xform => (reducer, coll, initValue) => {
  let state = {};
  state.value = typeof initState === 'function' ? initState() : initState;
  let step = enforceArgumentContract(process);
  let trans = (accum, input) => step(xform, reducer, accum, input, state);
  if (coll === undefined) {
    return trans; // return transducer
  } else if (typeof coll[Symbol.iterator] === 'function') {
    return unWrap(reduce(...[trans, coll, initValue].filter(exists))); 
  } else {
    throw NON_ITER;
  }
};

const take = factory((n, reducer, accum, input, state) => {
  if (state.value >= n) {
    return reduced(accum);
  } else {
    state.value += 1;
  }
  return reducer(accum, input);
}, () => 0);

如果你想看看这一切的实际效果，我不久前制作了一个小图书馆。虽然我忽略了 Cognitect 的互操作协议（我只是想了解概念），但我确实尝试根据 Strange Loop 和 Conj 的 Rich Hickey 的演讲尽可能准确地实现语义。