More transducers and reducing functions for Clojure(script)!
Transducers (in net.cgrand.xforms) can be classified in three groups: regular ones, higher-order ones
(which accept other transducers as arguments) and 1-item ones which emit only 1 item out no matter how many went in.
They generally only make sense in the context of a higher-order transducer.
- regular ones:
partition(1 arg),reductions,for,take-last,drop-last,windowandwindow-by-time - higher-order ones:
by-key,multiplex,transjuxt,partition(2+ args) - 1-item ones:
reduce,into,transjuxt,last,count,avg,sd,min,minimum,max,maximum,str
Reducing functions (in net.cgrand.xforms.rfs): min, minimum, max, maximum, str, str!, avg, sd, juxt and last.
Transducing contexts: transjuxt (for performing several transductions in a single pass), into, count.
Add this dependency to your project:
[net.cgrand/xforms "0.9.2"]=> (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 µsfor 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 µsYou 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 µsLike 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)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 a 1-item transducer 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]])(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).
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.
| fn | kvs in? | kvs out? |
|---|---|---|
| `for` | when first binding is a pair | when `body-expr` is a pair |
| `reduce` | when is `f` is a kvrf | no |
| 1-arg `into` (transducer) | when `to` is a map | no |
| 3-arg `into` (transducing context) | when `from` is a map | when `to` is a map |
| `by-key` (as a transducer) | when is `kfn` and `vfn` are unspecified or `nil` | when `pair` is `vector` or unspecified |
| `by-key` (as a transducing context on values) | no | no |
;; 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- Fix transients perf issue in Clojurescript
- Works with Clojurescript (even self-hosted).
- Added 2-arg arity to
x/countwhere it acts as a transducing context e.g.(x/count (filter odd?) (range 10)) - Preserve type hints in
x/for(and generally withkvrf).
- Added
x/reductions - Now if the first collection expression in
x/foris not a placeholder thenx/forworks likex/forbut returns an eduction and performs all iterations using reduce.
Copyright © 2015-2016 Christophe Grand
Distributed under the Eclipse Public License either version 1.0 or (at your option) any later version.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent