xforms

More transducers and reducing functions for Clojure(script)!

Transducers can be classified in three groups: regular ones, higher-order ones (which accept other transducers as arguments) and aggrerators (transdcuers which emit only 1 item out no matter how many went in). Aggregators generally only make sense in the context of a higher-order transducer.

In net.cgrand.xforms:

• regular ones: partition (1 arg), reductions, for, take-last, drop-last, sort, sort-by, wrap, window and window-by-time
• higher-order ones: by-key, into-by-key, multiplex, transjuxt, partition (2+ args)
• aggregators: reduce, into, without, transjuxt, last, count, avg, sd, min, minimum, max, maximum, str

In net.cgrand.xforms.io:

• sh to use any process as a transducer

Reducing functions

• in net.cgrand.xforms.rfs: min, minimum, max, maximum, str, str!, avg, sd, last and some.
• in net.cgrand.xforms.io: line-out and edn-out.

(in net.cgrand.xforms)

Transducing contexts:

• in net.cgrand.xforms: transjuxt (for performing several transductions in a single pass), iterator (clojure only), into, without, count, str (2 args) and some.
• in net.cgrand.xforms.io: line-out (3+ args) and edn-out (3+ args).
• in net.cgrand.xforms.nodejs.stream: transformer.

Reducible views (in net.cgrand.xforms.io): lines-in and edn-in.

Note: it should always be safe to update to the latest xforms version; short of bugfixes, breaking changes are avoided.

Usage

Add this dependency to your project:

[net.cgrand/xforms "0.16.0"]

=> (require '[net.cgrand.xforms :as x])

str and str! are two reducing functions to build Strings and StringBuilders in linear time.

=> (quick-bench (reduce str (range 256)))
             Execution time mean : 58,714946 µs
=> (quick-bench (reduce rf/str (range 256)))
             Execution time mean : 11,609631 µs

for is the transducing cousin of clojure.core/for:

=> (quick-bench (reduce + (for [i (range 128) j (range i)] (* i j))))
             Execution time mean : 514,932029 µs
=> (quick-bench (transduce (x/for [i % j (range i)] (* i j)) + 0 (range 128)))
             Execution time mean : 373,814060 µs

You can also use for like clojure.core/for: (x/for [i (range 128) j (range i)] (* i j)) expands to (eduction (x/for [i % j (range i)] (* i j)) (range 128)).

by-key and reduce are two new transducers. Here is an example usage:

;; reimplementing group-by
(defn my-group-by [kfn coll]
  (into {} (x/by-key kfn (x/reduce conj)) coll))

;; let's go transient!
(defn my-group-by [kfn coll]
  (into {} (x/by-key kfn (x/into [])) coll))

=> (quick-bench (group-by odd? (range 256)))
             Execution time mean : 29,356531 µs
=> (quick-bench (my-group-by odd? (range 256)))
             Execution time mean : 20,604297 µs

Like by-key, partition also takes a transducer as last argument to allow further computation on the partition.

=> (sequence (x/partition 4 (x/reduce +)) (range 16))
(6 22 38 54)

Padding is achieved as usual:

=> (sequence (x/partition 4 4 (repeat :pad) (x/into [])) (range 9))
([0 1 2 3] [4 5 6 7] [8 :pad :pad :pad])

avg is a transducer to compute the arithmetic mean. transjuxt is used to perform several transductions at once.

=> (into {} (x/by-key odd? (x/transjuxt [(x/reduce +) x/avg])) (range 256))
{false [16256 127], true [16384 128]}
=> (into {} (x/by-key odd? (x/transjuxt {:sum (x/reduce +) :mean x/avg :count x/count})) (range 256))
{false {:sum 16256, :mean 127, :count 128}, true {:sum 16384, :mean 128, :count 128}}

window is a new transducer to efficiently compute a windowed accumulator:

;; sum of last 3 items
=> (sequence (x/window 3 + -) (range 16))
(0 1 3 6 9 12 15 18 21 24 27 30 33 36 39 42)

=> (def nums (repeatedly 8 #(rand-int 42)))
#'user/nums
=> nums
(11 8 32 26 6 10 37 24)

;; avg of last 4 items
=> (sequence
     (x/window 4 x/avg #(x/avg %1 %2 -1))
     nums)
(11 19/2 17 77/4 18 37/2 79/4 77/4)

;; min of last 3 items
=> (sequence
     (x/window 3
       (fn
         ([] (sorted-set))
         ([s] (first s))
         ([s x] (conj s x)))
       disj)
     nums)
(11 8 8 8 6 6 6 10)

On Partitioning

Both by-key and partition takes a transducer as parameter. This transducer is used to further process each partition.

It's worth noting that all transformed outputs are subsequently interleaved. See:

=> (sequence (x/partition 2 1 identity) (range 8))
(0 1 1 2 2 3 3 4 4 5 5 6 6 7)
=> (sequence (x/by-key odd? identity) (range 8))
([false 0] [true 1] [false 2] [true 3] [false 4] [true 5] [false 6] [true 7])

That's why most of the time the last stage of the sub-transducer will be an aggregator like x/reduce or x/into:

=> (sequence (x/partition 2 1 (x/into [])) (range 8))
([0 1] [1 2] [2 3] [3 4] [4 5] [5 6] [6 7])
=> (sequence (x/by-key odd? (x/into [])) (range 8))
([false [0 2 4 6]] [true [1 3 5 7]])

Simple examples

(group-by kf coll) is (into {} (x/by-key kf (x/into []) coll)).

(plumbing/map-vals f m) is (into {} (x/by-key (map f)) m).

My faithful (reduce-by kf f init coll) is now (into {} (x/by-key kf (x/reduce f init))).

(frequencies coll) is (into {} (x/by-key identity x/count) coll).

On key-value pairs

Clojure reduce-kv is able to reduce key value pairs without allocating vectors or map entries: the key and value are passed as second and third arguments of the reducing function.

Xforms allows a reducing function to advertise its support for key value pairs (3-arg arity) by implementing the KvRfable protocol (in practice using the kvrf macro).

Several xforms transducers and transducing contexts leverage reduce-kv and kvrf. When these functions are used together, pairs can be transformed without being allocated.

;; plain old sequences
=> (let [m (zipmap (range 1e5) (range 1e5))]
     (crit/quick-bench
       (into {}
         (for [[k v] m]
           [k (inc v)]))))
Evaluation count : 12 in 6 samples of 2 calls.
             Execution time mean : 55,150081 ms
    Execution time std-deviation : 1,397185 ms

;; x/for but pairs are allocated (because of into) 
=> (let [m (zipmap (range 1e5) (range 1e5))]
     (crit/quick-bench
       (into {}
         (x/for [[k v] _]
           [k (inc v)])
         m)))
Evaluation count : 18 in 6 samples of 3 calls.
             Execution time mean : 39,119387 ms
    Execution time std-deviation : 1,456902 ms
    
;; x/for but no pairs are allocated (thanks to x/into) 
=> (let [m (zipmap (range 1e5) (range 1e5))]
     (crit/quick-bench (x/into {}
               (x/for [[k v] %]
                 [k (inc v)])
               m)))
Evaluation count : 24 in 6 samples of 4 calls.
             Execution time mean : 24,276790 ms
    Execution time std-deviation : 364,932996 µs

Changelog

0.9.5

• Short (up to 4) literal collections (or literal collections with :unroll metadata) in collection positions in x/for are unrolled. This means that the collection is not allocated. If it's a collection of pairs (e.g. maps), pairs themselves won't be allocated.

0.9.4

• Add x/into-by-key short hand

0.7.2

• Fix transients perf issue in Clojurescript

0.7.1

• Works with Clojurescript (even self-hosted).

0.7.0

• Added 2-arg arity to x/count where it acts as a transducing context e.g. (x/count (filter odd?) (range 10))
• Preserve type hints in x/for (and generally with kvrf).

0.6.0

• Added x/reductions
• Now if the first collection expression in x/for is not a placeholder then x/for works like x/for but returns an eduction and performs all iterations using reduce.

License

Copyright © 2015-2016 Christophe Grand

Distributed under the Eclipse Public License either version 1.0 or (at your option) any later version.