1 (* Title: HOL/Tools/Sledgehammer/sledgehammer_mash.ML
2 Author: Jasmin Blanchette, TU Muenchen
3 Author: Cezary Kaliszyk, University of Innsbruck
5 Sledgehammer's machine-learning-based relevance filter (MaSh).
8 signature SLEDGEHAMMER_MASH =
10 type stature = ATP_Problem_Generate.stature
11 type raw_fact = Sledgehammer_Fact.raw_fact
12 type fact = Sledgehammer_Fact.fact
13 type fact_override = Sledgehammer_Fact.fact_override
14 type params = Sledgehammer_Prover.params
15 type prover_result = Sledgehammer_Prover.prover_result
17 val trace : bool Config.T
18 val duplicates : bool Config.T
26 val learn_isarN : string
27 val learn_proverN : string
28 val relearn_isarN : string
29 val relearn_proverN : string
30 val fact_filters : string list
31 val encode_str : string -> string
32 val encode_strs : string list -> string
33 val decode_str : string -> string
34 val decode_strs : string -> string list
36 datatype mash_algorithm =
43 val is_mash_enabled : unit -> bool
44 val the_mash_algorithm : unit -> mash_algorithm
45 val str_of_mash_algorithm : mash_algorithm -> string
47 val mesh_facts : ('a * 'a -> bool) -> int -> (real * (('a * real) list * 'a list)) list -> 'a list
48 val nickname_of_thm : thm -> string
49 val find_suggested_facts : Proof.context -> ('b * thm) list -> string list -> ('b * thm) list
50 val crude_thm_ord : thm * thm -> order
51 val thm_less : thm * thm -> bool
52 val goal_of_thm : theory -> thm -> thm
53 val run_prover_for_mash : Proof.context -> params -> string -> string -> fact list -> thm ->
55 val features_of : Proof.context -> theory -> stature -> term list -> string list
56 val trim_dependencies : string list -> string list option
57 val isar_dependencies_of : string Symtab.table * string Symtab.table -> thm -> string list option
58 val prover_dependencies_of : Proof.context -> params -> string -> int -> raw_fact list ->
59 string Symtab.table * string Symtab.table -> thm -> bool * string list
60 val attach_parents_to_facts : ('a * thm) list -> ('a * thm) list ->
61 (string list * ('a * thm)) list
62 val num_extra_feature_facts : int
63 val extra_feature_factor : real
64 val weight_facts_smoothly : 'a list -> ('a * real) list
65 val weight_facts_steeply : 'a list -> ('a * real) list
66 val find_mash_suggestions : Proof.context -> int -> string list -> ('a * thm) list ->
67 ('a * thm) list -> ('a * thm) list -> ('a * thm) list * ('a * thm) list
68 val mash_suggested_facts : Proof.context -> theory -> params -> int -> term list -> term ->
69 raw_fact list -> fact list * fact list
71 val mash_unlearn : unit -> unit
72 val mash_learn_proof : Proof.context -> params -> term -> ('a * thm) list -> thm list -> unit
73 val mash_learn_facts : Proof.context -> params -> string -> int -> bool -> Time.time ->
74 raw_fact list -> string
75 val mash_learn : Proof.context -> params -> fact_override -> thm list -> bool -> unit
76 val mash_can_suggest_facts : Proof.context -> bool
78 val generous_max_suggestions : int -> int
79 val mepo_weight : real
80 val mash_weight : real
81 val relevant_facts : Proof.context -> params -> string -> int -> fact_override -> term list ->
82 term -> raw_fact list -> (string * fact list) list
83 val kill_learners : unit -> unit
84 val running_learners : unit -> unit
87 structure Sledgehammer_MaSh : SLEDGEHAMMER_MASH =
91 open ATP_Problem_Generate
92 open Sledgehammer_Util
93 open Sledgehammer_Fact
94 open Sledgehammer_Prover
95 open Sledgehammer_Prover_Minimize
96 open Sledgehammer_MePo
98 val trace = Attrib.setup_config_bool @{binding sledgehammer_mash_trace} (K false)
99 val duplicates = Attrib.setup_config_bool @{binding sledgehammer_fact_duplicates} (K false)
101 fun trace_msg ctxt msg = if Config.get ctxt trace then tracing (msg ()) else ()
103 fun gen_eq_thm ctxt = if Config.get ctxt duplicates then Thm.eq_thm_strict else Thm.eq_thm_prop
113 val fact_filters = [meshN, mepoN, mashN]
115 val unlearnN = "unlearn"
116 val learn_isarN = "learn_isar"
117 val learn_proverN = "learn_prover"
118 val relearn_isarN = "relearn_isar"
119 val relearn_proverN = "relearn_prover"
121 fun map_array_at ary f i = Array.update (ary, i, f (Array.sub (ary, i)))
123 type xtab = int * int Symtab.table
125 val empty_xtab = (0, Symtab.empty)
127 fun add_to_xtab key (next, tab) = (next + 1, Symtab.update_new (key, next) tab)
128 fun maybe_add_to_xtab key = perhaps (try (add_to_xtab key))
130 fun state_file () = Path.expand (Path.explode "$ISABELLE_HOME_USER/mash_state")
131 val remove_state_file = try File.rm o state_file
133 datatype mash_algorithm =
140 (* TODO: eliminate "MASH" environment variable after Isabelle2014 release *)
141 fun mash_algorithm () =
142 let val flag1 = Options.default_string @{system_option MaSh} in
143 (case if flag1 <> "none" (* default *) then flag1 else getenv "MASH" of
144 "yes" => SOME MaSh_NB_kNN
145 | "sml" => SOME MaSh_NB_kNN
146 | "nb" => SOME MaSh_NB
147 | "knn" => SOME MaSh_kNN
148 | "nb_knn" => SOME MaSh_NB_kNN
149 | "nb_ext" => SOME MaSh_NB_Ext
150 | "knn_ext" => SOME MaSh_kNN_Ext
152 | algorithm => (warning ("Unknown MaSh algorithm: " ^ quote algorithm ^ "."); NONE))
155 val is_mash_enabled = is_some o mash_algorithm
156 val the_mash_algorithm = the_default MaSh_NB_kNN o mash_algorithm
158 fun str_of_mash_algorithm MaSh_NB = "nb"
159 | str_of_mash_algorithm MaSh_kNN = "knn"
160 | str_of_mash_algorithm MaSh_NB_kNN = "nb_knn"
161 | str_of_mash_algorithm MaSh_NB_Ext = "nb_ext"
162 | str_of_mash_algorithm MaSh_kNN_Ext = "knn_ext"
164 fun scaled_avg [] = 0
165 | scaled_avg xs = Real.ceil (100000000.0 * fold (curry (op +)) xs 0.0) div length xs
168 | avg xs = fold (curry (op +)) xs 0.0 / Real.fromInt (length xs)
170 fun normalize_scores _ [] = []
171 | normalize_scores max_facts xs =
172 map (apsnd (curry Real.* (1.0 / avg (map snd (take max_facts xs))))) xs
174 fun mesh_facts fact_eq max_facts [(_, (sels, unks))] =
175 distinct fact_eq (map fst (take max_facts sels) @ take (max_facts - length sels) unks)
176 | mesh_facts fact_eq max_facts mess =
178 val mess = mess |> map (apsnd (apfst (normalize_scores max_facts)))
180 fun score_in fact (global_weight, (sels, unks)) =
181 let val score_at = try (nth sels) #> Option.map (fn (_, score) => global_weight * score) in
182 (case find_index (curry fact_eq fact o fst) sels of
183 ~1 => if member fact_eq unks fact then NONE else SOME 0.0
184 | rank => score_at rank)
187 fun weight_of fact = mess |> map_filter (score_in fact) |> scaled_avg
189 fold (union fact_eq o map fst o take max_facts o fst o snd) mess []
190 |> map (`weight_of) |> sort (int_ord o apply2 fst o swap)
191 |> map snd |> take max_facts
194 fun smooth_weight_of_fact rank = Math.pow (1.3, 15.5 - 0.2 * Real.fromInt rank) + 15.0 (* FUDGE *)
195 fun steep_weight_of_fact rank = Math.pow (0.62, log2 (Real.fromInt (rank + 1))) (* FUDGE *)
197 fun weight_facts_smoothly facts = facts ~~ map smooth_weight_of_fact (0 upto length facts - 1)
198 fun weight_facts_steeply facts = facts ~~ map steep_weight_of_fact (0 upto length facts - 1)
200 fun sort_array_suffix cmp needed a =
202 exception BOTTOM of int
204 val al = Array.length a
207 let val i31 = i + i + i + 1 in
209 let val x = Unsynchronized.ref i31 in
210 if cmp (Array.sub (a, i31), Array.sub (a, i31 + 1)) = LESS then x := i31 + 1 else ();
211 if cmp (Array.sub (a, !x), Array.sub (a, i31 + 2)) = LESS then x := i31 + 2 else ();
215 if i31 + 1 < l andalso cmp (Array.sub (a, i31), Array.sub (a, i31 + 1)) = LESS
216 then i31 + 1 else if i31 < l then i31 else raise BOTTOM i
219 fun trickledown l i e =
220 let val j = maxson l i in
221 if cmp (Array.sub (a, j), e) = GREATER then
222 (Array.update (a, i, Array.sub (a, j)); trickledown l j e)
224 Array.update (a, i, e)
227 fun trickle l i e = trickledown l i e handle BOTTOM i => Array.update (a, i, e)
230 let val j = maxson l i in
231 Array.update (a, i, Array.sub (a, j));
235 fun bubble l i = bubbledown l i handle BOTTOM i => i
238 let val father = (i - 1) div 3 in
239 if cmp (Array.sub (a, father), e) = LESS then
240 (Array.update (a, i, Array.sub (a, father));
241 if father > 0 then trickleup father e else Array.update (a, 0, e))
243 Array.update (a, i, e)
246 fun for i = if i < 0 then () else (trickle al i (Array.sub (a, i)); for (i - 1))
249 if i < Integer.max 2 (al - needed) then
252 let val e = Array.sub (a, i) in
253 Array.update (a, i, Array.sub (a, 0));
254 trickleup (bubble i 0) e;
258 for (((al + 1) div 3) - 1);
261 let val e = Array.sub (a, 1) in
262 Array.update (a, 1, Array.sub (a, 0));
263 Array.update (a, 0, e)
269 fun rev_sort_list_prefix cmp needed xs =
270 let val ary = Array.fromList xs in
271 sort_array_suffix cmp needed ary;
272 Array.foldl (op ::) [] ary
276 (*** Isabelle-agnostic machine learning ***)
281 fun select_visible_facts (big_number : real) recommends =
283 let val (j, ov) = Array.sub (recommends, at) in
284 Array.update (recommends, at, (j, big_number + ov))
287 fun wider_array_of_vector init vec =
288 let val ary = Array.array init in
289 Array.copyVec {src = vec, dst = ary, di = 0};
293 val nb_def_prior_weight = 1000 (* FUDGE *)
295 fun learn_facts (tfreq0, sfreq0, dffreq0) num_facts0 num_facts num_feats depss featss =
297 val tfreq = wider_array_of_vector (num_facts, 0) tfreq0
298 val sfreq = wider_array_of_vector (num_facts, Inttab.empty) sfreq0
299 val dffreq = wider_array_of_vector (num_feats, 0) dffreq0
301 fun learn_one th feats deps =
303 fun add_th weight t =
305 val im = Array.sub (sfreq, t)
306 fun fold_fn s = Inttab.map_default (s, 0) (Integer.add weight)
308 map_array_at tfreq (Integer.add weight) t;
309 Array.update (sfreq, t, fold fold_fn feats im)
312 val add_sym = map_array_at dffreq (Integer.add 1)
314 add_th nb_def_prior_weight th;
315 List.app (add_th 1) deps;
316 List.app add_sym feats
320 if i = num_facts then ()
321 else (learn_one i (Vector.sub (featss, i)) (Vector.sub (depss, i)); for (i + 1))
324 (Array.vector tfreq, Array.vector sfreq, Array.vector dffreq)
327 fun naive_bayes (tfreq, sfreq, dffreq) num_facts max_suggs visible_facts goal_feats =
329 val tau = 0.2 (* FUDGE *)
330 val pos_weight = 5.0 (* FUDGE *)
331 val def_val = ~18.0 (* FUDGE *)
332 val init_val = 30.0 (* FUDGE *)
334 val ln_afreq = Math.ln (Real.fromInt num_facts)
335 val idf = Vector.map (fn i => ln_afreq - Math.ln (Real.fromInt i)) dffreq
337 fun tfidf feat = Vector.sub (idf, feat)
339 fun log_posterior i =
341 val tfreq = Real.fromInt (Vector.sub (tfreq, i))
343 fun add_feat (f, fw0) (res, sfh) =
344 (case Inttab.lookup sfh f of
346 (res + fw0 * tfidf f * Math.ln (pos_weight * Real.fromInt sf / tfreq),
348 | NONE => (res + fw0 * tfidf f * def_val, sfh))
350 val (res, sfh) = fold add_feat goal_feats (init_val * Math.ln tfreq, Vector.sub (sfreq, i))
352 fun fold_sfh (f, sf) sow =
353 sow + tfidf f * Math.ln (1.0 - Real.fromInt (sf - 1) / tfreq)
355 val sum_of_weights = Inttab.fold fold_sfh sfh 0.0
357 res + tau * sum_of_weights
360 val posterior = Array.tabulate (num_facts, (fn j => (j, log_posterior j)))
363 if at = num_facts then acc else ret (at + 1) (Array.sub (posterior, at) :: acc)
365 select_visible_facts 100000.0 posterior visible_facts;
366 sort_array_suffix (Real.compare o apply2 snd) max_suggs posterior;
367 ret (Integer.max 0 (num_facts - max_suggs)) []
372 fun k_nearest_neighbors dffreq num_facts num_feats depss featss max_suggs visible_facts goal_feats =
374 exception EXIT of unit
376 val ln_afreq = Math.ln (Real.fromInt num_facts)
377 fun tfidf feat = ln_afreq - Math.ln (Real.fromInt (Vector.sub (dffreq, feat)))
379 val overlaps_sqr = Array.tabulate (num_facts, rpair 0.0)
381 val feat_facts = Array.array (num_feats, [])
382 val _ = Vector.foldl (fn (feats, fact) =>
383 (List.app (map_array_at feat_facts (cons fact)) feats; fact + 1)) 0 featss
385 fun do_feat (s, sw0) =
387 val sw = sw0 * tfidf s
388 val w6 = Math.pow (sw, 6.0 (* FUDGE *))
391 let val (_, ov) = Array.sub (overlaps_sqr, j) in
392 Array.update (overlaps_sqr, j, (j, w6 + ov))
395 List.app inc_overlap (Array.sub (feat_facts, s))
398 val _ = List.app do_feat goal_feats
399 val _ = sort_array_suffix (Real.compare o apply2 snd) num_facts overlaps_sqr
400 val no_recommends = Unsynchronized.ref 0
401 val recommends = Array.tabulate (num_facts, rpair 0.0)
402 val age = Unsynchronized.ref 500000000.0
404 fun inc_recommend v j =
405 let val (_, ov) = Array.sub (recommends, j) in
407 (no_recommends := !no_recommends + 1; Array.update (recommends, j, (j, !age + ov)))
409 Array.update (recommends, j, (j, v + ov))
412 val k = Unsynchronized.ref 0
414 if k >= num_facts then
418 val deps_factor = 2.7 (* FUDGE *)
419 val (j, o2) = Array.sub (overlaps_sqr, num_facts - k - 1)
420 val _ = inc_recommend o2 j
421 val ds = Vector.sub (depss, j)
422 val l = Real.fromInt (length ds)
424 List.app (inc_recommend (deps_factor * o2 / l)) ds
428 if !k = initial_k + 1 then () else (do_k (!k); k := !k + 1; while1 ())
432 if !no_recommends >= max_suggs then ()
433 else (do_k (!k); k := !k + 1; age := !age - 10000.0; while2 ())
437 if at = num_facts then acc else ret (Array.sub (recommends, at) :: acc) (at + 1)
441 select_visible_facts 1000000000.0 recommends visible_facts;
442 sort_array_suffix (Real.compare o apply2 snd) max_suggs recommends;
443 ret [] (Integer.max 0 (num_facts - max_suggs))
447 fun external_tool tool max_suggs learns goal_feats =
449 val ser = string_of_int (serial ()) (* poor person's attempt at thread-safety *)
450 val ocs = TextIO.openOut ("adv_syms" ^ ser)
451 val ocd = TextIO.openOut ("adv_deps" ^ ser)
452 val ocq = TextIO.openOut ("adv_seq" ^ ser)
453 val occ = TextIO.openOut ("adv_conj" ^ ser)
455 fun os oc s = TextIO.output (oc, s)
459 | ol oc f sep (h :: t) = (f h; os oc sep; ol oc f sep t)
461 fun do_learn (name, feats, deps) =
462 (os ocs name; os ocs ":"; ol ocs (os ocs o quote) ", " feats; os ocs "\n";
463 os ocd name; os ocd ":"; ol ocd (os ocd) " " deps; os ocd "\n"; os ocq name; os ocq "\n")
468 "~/misc/" ^ tool ^ " adv_syms" ^ ser ^ " adv_deps" ^ ser ^ " " ^ string_of_int no ^
469 " adv_seq" ^ ser ^ " < adv_conj" ^ ser
471 fst (Isabelle_System.bash_output cmd)
473 |> filter_out (curry (op =) "")
476 (List.app do_learn learns; ol occ (os occ o quote) ", " (map fst goal_feats);
477 TextIO.closeOut ocs; TextIO.closeOut ocd; TextIO.closeOut ocq; TextIO.closeOut occ;
481 fun k_nearest_neighbors_ext max_suggs =
482 external_tool ("newknn/knn" ^ " " ^ string_of_int initial_k) max_suggs
483 fun naive_bayes_ext max_suggs = external_tool "predict/nbayes" max_suggs
485 fun query_external ctxt algorithm max_suggs learns goal_feats =
486 (trace_msg ctxt (fn () => "MaSh query external " ^ commas (map fst goal_feats));
488 MaSh_NB_Ext => naive_bayes_ext max_suggs learns goal_feats
489 | MaSh_kNN_Ext => k_nearest_neighbors_ext max_suggs learns goal_feats))
491 fun query_internal ctxt algorithm num_facts num_feats (fact_names, featss, depss)
492 (freqs as (_, _, dffreq)) visible_facts max_suggs goal_feats int_goal_feats =
495 naive_bayes freqs num_facts max_suggs visible_facts int_goal_feats
498 k_nearest_neighbors dffreq num_facts num_feats depss featss max_suggs visible_facts
502 (trace_msg ctxt (fn () => "MaSh query internal " ^ commas (map fst goal_feats) ^ " from {" ^
503 elide_string 1000 (space_implode " " (Vector.foldr (op ::) [] fact_names)) ^ "}");
508 mesh_facts (op =) max_suggs
509 [(0.5 (* FUDGE *), (weight_facts_steeply (nb ()), [])),
510 (0.5 (* FUDGE *), (weight_facts_steeply (knn ()), []))])
511 |> map (curry Vector.sub fact_names))
517 (*** Persistent, stringly-typed state ***)
520 if Char.isAlphaNum c orelse c = #"_" orelse c = #"." orelse c = #"(" orelse c = #")" orelse
524 (* fixed width, in case more digits follow *)
525 "%" ^ stringN_of_int 3 (Char.ord c)
527 fun unmeta_chars accum [] = String.implode (rev accum)
528 | unmeta_chars accum (#"%" :: d1 :: d2 :: d3 :: cs) =
529 (case Int.fromString (String.implode [d1, d2, d3]) of
530 SOME n => unmeta_chars (Char.chr n :: accum) cs
531 | NONE => "" (* error *))
532 | unmeta_chars _ (#"%" :: _) = "" (* error *)
533 | unmeta_chars accum (c :: cs) = unmeta_chars (c :: accum) cs
535 val encode_str = String.translate meta_char
536 val decode_str = String.explode #> unmeta_chars []
538 val encode_strs = map encode_str #> space_implode " "
539 val decode_strs = space_explode " " #> map decode_str
541 datatype proof_kind = Isar_Proof | Automatic_Proof | Isar_Proof_wegen_Prover_Flop
543 fun str_of_proof_kind Isar_Proof = "i"
544 | str_of_proof_kind Automatic_Proof = "a"
545 | str_of_proof_kind Isar_Proof_wegen_Prover_Flop = "x"
547 fun proof_kind_of_str "a" = Automatic_Proof
548 | proof_kind_of_str "x" = Isar_Proof_wegen_Prover_Flop
549 | proof_kind_of_str _ (* "i" *) = Isar_Proof
551 fun add_edge_to name parent =
552 Graph.default_node (parent, (Isar_Proof, [], []))
553 #> Graph.add_edge (parent, name)
555 fun add_node kind name parents feats deps (accum as (access_G, (fact_xtab, feat_xtab), learns)) =
556 let val fact_xtab' = add_to_xtab name fact_xtab in
557 ((Graph.new_node (name, (kind, feats, deps)) access_G
558 handle Graph.DUP _ => Graph.map_node name (K (kind, feats, deps)) access_G)
559 |> fold (add_edge_to name) parents,
560 (fact_xtab', fold maybe_add_to_xtab feats feat_xtab),
561 (name, feats, deps) :: learns)
563 handle Symtab.DUP _ => accum (* robustness (in case the state file violates the invariant) *)
565 fun try_graph ctxt when def f =
568 Graph.CYCLES (cycle :: _) =>
569 (trace_msg ctxt (fn () => "Cycle involving " ^ commas cycle ^ " when " ^ when); def)
571 (trace_msg ctxt (fn () => "Duplicate fact " ^ quote name ^ " when " ^ when); def)
572 | Graph.UNDEF name =>
573 (trace_msg ctxt (fn () => "Unknown fact " ^ quote name ^ " when " ^ when); def)
575 if Exn.is_interrupt exn then
578 (trace_msg ctxt (fn () => "Internal error when " ^ when ^ ":\n" ^ Runtime.exn_message exn);
582 string_of_int (length (Graph.keys G)) ^ " node(s), " ^
583 string_of_int (fold (Integer.add o length o snd) (Graph.dest G) 0) ^ " edge(s), " ^
584 string_of_int (length (Graph.maximals G)) ^ " maximal"
586 type ffds = string vector * int list vector * int list vector
587 type freqs = int vector * int Inttab.table vector * int vector
590 {access_G : (proof_kind * string list * string list) Graph.T,
594 dirty_facts : string list option}
596 val empty_xtabs = (empty_xtab, empty_xtab)
597 val empty_ffds = (Vector.fromList [], Vector.fromList [], Vector.fromList []) : ffds
598 val empty_freqs = (Vector.fromList [], Vector.fromList [], Vector.fromList []) : freqs
601 {access_G = Graph.empty,
605 dirty_facts = SOME []} : mash_state
607 fun recompute_ffds_freqs_from_learns (learns : (string * string list * string list) list)
608 ((num_facts, fact_tab), (num_feats, feat_tab)) num_facts0 (fact_names0, featss0, depss0) freqs0 =
610 val fact_names = Vector.concat [fact_names0, Vector.fromList (map #1 learns)]
611 val featss = Vector.concat [featss0,
612 Vector.fromList (map (map_filter (Symtab.lookup feat_tab) o #2) learns)]
613 val depss = Vector.concat [depss0,
614 Vector.fromList (map (map_filter (Symtab.lookup fact_tab) o #3) learns)]
616 ((fact_names, featss, depss),
617 MaSh.learn_facts freqs0 num_facts0 num_facts num_feats depss featss)
620 fun reorder_learns (num_facts, fact_tab) learns =
621 let val ary = Array.array (num_facts, ("", [], [])) in
622 List.app (fn learn as (fact, _, _) =>
623 Array.update (ary, the (Symtab.lookup fact_tab fact), learn))
625 Array.foldr (op ::) [] ary
628 fun recompute_ffds_freqs_from_access_G access_G (xtabs as (fact_xtab, _)) =
631 Graph.schedule (fn _ => fn (fact, (_, feats, deps)) => (fact, feats, deps)) access_G
632 |> reorder_learns fact_xtab
634 recompute_ffds_freqs_from_learns learns xtabs 0 empty_ffds empty_freqs
639 val version = "*** MaSh version 20140625 ***"
641 exception FILE_VERSION_TOO_NEW of unit
643 fun extract_node line =
644 (case space_explode ":" line of
646 (case (space_explode " " head, map (unprefix " ") (space_explode ";" tail)) of
647 ([kind, name], [parents, feats, deps]) =>
648 SOME (proof_kind_of_str kind, decode_str name, decode_strs parents, decode_strs feats,
653 fun load_state ctxt (time_state as (memory_time, _)) =
654 let val path = state_file () in
655 (case try OS.FileSys.modTime (Path.implode path) of
658 if Time.>= (memory_time, disk_time) then
662 (case try File.read_lines path of
663 SOME (version' :: node_lines) =>
665 fun extract_line_and_add_node line =
666 (case extract_node line of
667 NONE => I (* should not happen *)
668 | SOME (kind, name, parents, feats, deps) => add_node kind name parents feats deps)
670 val empty_G_etc = (Graph.empty, empty_xtabs, [])
672 val (access_G, xtabs, rev_learns) =
673 (case string_ord (version', version) of
675 try_graph ctxt "loading state" empty_G_etc
676 (fn () => fold extract_line_and_add_node node_lines empty_G_etc)
677 | LESS => (remove_state_file (); empty_G_etc) (* cannot parse old file *)
678 | GREATER => raise FILE_VERSION_TOO_NEW ())
681 recompute_ffds_freqs_from_learns (rev rev_learns) xtabs 0 empty_ffds empty_freqs
683 trace_msg ctxt (fn () => "Loaded fact graph (" ^ graph_info access_G ^ ")");
684 {access_G = access_G, xtabs = xtabs, ffds = ffds, freqs = freqs, dirty_facts = SOME []}
686 | _ => empty_state)))
689 fun str_of_entry (kind, name, parents, feats, deps) =
690 str_of_proof_kind kind ^ " " ^ encode_str name ^ ": " ^ encode_strs parents ^ "; " ^
691 encode_strs feats ^ "; " ^ encode_strs deps ^ "\n"
693 fun save_state _ (time_state as (_, {dirty_facts = SOME [], ...})) = time_state
694 | save_state ctxt (memory_time, {access_G, xtabs, ffds, freqs, dirty_facts}) =
696 fun append_entry (name, ((kind, feats, deps), (parents, _))) =
697 cons (kind, name, Graph.Keys.dest parents, feats, deps)
699 val path = state_file ()
701 (case try OS.FileSys.modTime (Path.implode path) of
703 | SOME disk_time => if Time.<= (disk_time, memory_time) then dirty_facts else NONE)
704 val (banner, entries) =
705 (case dirty_facts' of
706 SOME names => (NONE, fold (append_entry o Graph.get_entry access_G) names [])
707 | NONE => (SOME (version ^ "\n"), Graph.fold append_entry access_G []))
709 (case banner of SOME s => File.write path s | NONE => ();
710 entries |> chunk_list 500 |> List.app (File.append path o implode o map str_of_entry))
711 handle IO.Io _ => ();
712 trace_msg ctxt (fn () =>
713 "Saved fact graph (" ^ graph_info access_G ^
715 SOME dirty_facts => "; " ^ string_of_int (length dirty_facts) ^ " dirty fact(s)"
718 {access_G = access_G, xtabs = xtabs, ffds = ffds, freqs = freqs, dirty_facts = SOME []})
721 val global_state = Synchronized.var "Sledgehammer_MaSh.global_state" (Time.zeroTime, empty_state)
725 fun map_state ctxt f =
726 Synchronized.change global_state (load_state ctxt ##> f #> save_state ctxt)
727 handle FILE_VERSION_TOO_NEW () => ()
729 fun peek_state ctxt =
730 Synchronized.change_result global_state (perhaps (try (load_state ctxt)) #> `snd)
733 Synchronized.change global_state (fn _ => (remove_state_file (); (Time.zeroTime, empty_state)))
738 (*** Isabelle helpers ***)
740 fun elided_backquote_thm threshold th =
741 elide_string threshold (backquote_thm (Proof_Context.init_global (Thm.theory_of_thm th)) th)
743 val thy_name_of_thm = Context.theory_name o Thm.theory_of_thm
745 fun nickname_of_thm th =
746 if Thm.has_name_hint th then
747 let val hint = Thm.get_name_hint th in
748 (* There must be a better way to detect local facts. *)
749 (case Long_Name.dest_local hint of
750 SOME suf => Long_Name.implode [thy_name_of_thm th, suf, elided_backquote_thm 50 th]
754 elided_backquote_thm 200 th
756 fun find_suggested_facts ctxt facts =
758 fun add (fact as (_, th)) = Symtab.default (nickname_of_thm th, fact)
759 val tab = fold add facts Symtab.empty
761 Symtab.lookup tab nick
762 |> tap (fn NONE => trace_msg ctxt (fn () => "Cannot find " ^ quote nick) | _ => ())
763 in map_filter lookup end
765 fun free_feature_of s = "f" ^ s
766 fun thy_feature_of s = "y" ^ s
767 fun type_feature_of s = "t" ^ s
768 fun class_feature_of s = "s" ^ s
769 val local_feature = "local"
771 fun crude_theory_ord p =
772 if Theory.subthy p then
773 if Theory.eq_thy p then EQUAL else LESS
774 else if Theory.subthy (swap p) then
777 (case int_ord (apply2 (length o Theory.ancestors_of) p) of
778 EQUAL => string_ord (apply2 Context.theory_name p)
781 fun crude_thm_ord p =
782 (case crude_theory_ord (apply2 Thm.theory_of_thm p) of
784 (* The hack below is necessary because of odd dependencies that are not reflected in the theory
786 let val q = apply2 nickname_of_thm p in
787 (* Hack to put "xxx_def" before "xxxI" and "xxxE" *)
788 (case bool_ord (apply2 (String.isSuffix "_def") (swap q)) of
789 EQUAL => string_ord q
794 val thm_less_eq = Theory.subthy o apply2 Thm.theory_of_thm
795 fun thm_less p = thm_less_eq p andalso not (thm_less_eq (swap p))
797 val freezeT = Type.legacy_freeze_type
799 fun freeze (t $ u) = freeze t $ freeze u
800 | freeze (Abs (s, T, t)) = Abs (s, freezeT T, freeze t)
801 | freeze (Var ((s, _), T)) = Free (s, freezeT T)
802 | freeze (Const (s, T)) = Const (s, freezeT T)
803 | freeze (Free (s, T)) = Free (s, freezeT T)
806 fun goal_of_thm thy = Thm.prop_of #> freeze #> Thm.global_cterm_of thy #> Goal.init
808 fun run_prover_for_mash ctxt params prover goal_name facts goal =
811 {comment = "Goal: " ^ goal_name, state = Proof.init ctxt, goal = goal, subgoal = 1,
812 subgoal_count = 1, factss = [("", facts)]}
814 get_minimizing_prover ctxt MaSh (K ()) prover params problem
817 val bad_types = [@{type_name prop}, @{type_name bool}, @{type_name fun}]
819 val pat_tvar_prefix = "_"
820 val pat_var_prefix = "_"
822 (* try "Long_Name.base_name" for shorter names *)
823 fun massage_long_name s = s
825 val crude_str_of_sort = space_implode ":" o map massage_long_name o subtract (op =) @{sort type}
827 fun crude_str_of_typ (Type (s, [])) = massage_long_name s
828 | crude_str_of_typ (Type (s, Ts)) = massage_long_name s ^ implode (map crude_str_of_typ Ts)
829 | crude_str_of_typ (TFree (_, S)) = crude_str_of_sort S
830 | crude_str_of_typ (TVar (_, S)) = crude_str_of_sort S
832 fun maybe_singleton_str _ "" = []
833 | maybe_singleton_str pref s = [pref ^ s]
835 val max_pat_breadth = 10 (* FUDGE *)
837 fun term_features_of ctxt thy_name term_max_depth type_max_depth ts =
839 val thy = Proof_Context.theory_of ctxt
841 val fixes = map snd (Variable.dest_fixes ctxt)
842 val classes = Sign.classes_of thy
844 fun add_classes @{sort type} = I
846 fold (`(Sorts.super_classes classes)
848 #> subtract (op =) @{sort type} #> map massage_long_name
849 #> map class_feature_of
852 fun pattify_type 0 _ = []
853 | pattify_type _ (Type (s, [])) =
854 if member (op =) bad_types s then [] else [massage_long_name s]
855 | pattify_type depth (Type (s, U :: Ts)) =
858 val ps = take max_pat_breadth (pattify_type depth T)
859 val qs = take max_pat_breadth ("" :: pattify_type (depth - 1) U)
861 map_product (fn p => fn "" => p | q => p ^ "(" ^ q ^ ")") ps qs
863 | pattify_type _ (TFree (_, S)) = maybe_singleton_str pat_tvar_prefix (crude_str_of_sort S)
864 | pattify_type _ (TVar (_, S)) = maybe_singleton_str pat_tvar_prefix (crude_str_of_sort S)
866 fun add_type_pat depth T =
867 union (op =) (map type_feature_of (pattify_type depth T))
869 fun add_type_pats 0 _ = I
870 | add_type_pats depth t = add_type_pat depth t #> add_type_pats (depth - 1) t
873 add_type_pats type_max_depth T
874 #> fold_atyps_sorts (add_classes o snd) T
876 fun add_subtypes (T as Type (_, Ts)) = add_type T #> fold add_subtypes Ts
877 | add_subtypes T = add_type T
879 fun pattify_term _ 0 _ = []
880 | pattify_term _ _ (Const (s, _)) =
881 if is_widely_irrelevant_const s then [] else [massage_long_name s]
882 | pattify_term _ _ (Free (s, T)) =
883 maybe_singleton_str pat_var_prefix (crude_str_of_typ T)
884 |> (if member (op =) fixes s then
885 cons (free_feature_of (massage_long_name (Long_Name.append thy_name s)))
888 | pattify_term _ _ (Var (_, T)) = maybe_singleton_str pat_var_prefix (crude_str_of_typ T)
889 | pattify_term Ts _ (Bound j) =
890 maybe_singleton_str pat_var_prefix (crude_str_of_typ (nth Ts j))
891 | pattify_term Ts depth (t $ u) =
893 val ps = take max_pat_breadth (pattify_term Ts depth t)
894 val qs = take max_pat_breadth ("" :: pattify_term Ts (depth - 1) u)
896 map_product (fn p => fn "" => p | q => p ^ "(" ^ q ^ ")") ps qs
898 | pattify_term _ _ _ = []
900 fun add_term_pat Ts = union (op =) oo pattify_term Ts
902 fun add_term_pats _ 0 _ = I
903 | add_term_pats Ts depth t = add_term_pat Ts depth t #> add_term_pats Ts (depth - 1) t
905 fun add_term Ts = add_term_pats Ts term_max_depth
907 fun add_subterms Ts t =
908 (case strip_comb t of
909 (Const (s, T), args) =>
910 (not (is_widely_irrelevant_const s) ? add_term Ts t)
911 #> add_subtypes T #> fold (add_subterms Ts) args
914 Free (_, T) => add_term Ts t #> add_subtypes T
915 | Var (_, T) => add_subtypes T
916 | Abs (_, T, body) => add_subtypes T #> add_subterms (T :: Ts) body
918 #> fold (add_subterms Ts) args)
920 fold (add_subterms []) ts []
923 val term_max_depth = 2
924 val type_max_depth = 1
926 (* TODO: Generate type classes for types? *)
927 fun features_of ctxt thy (scope, _) ts =
928 let val thy_name = Context.theory_name thy in
929 thy_feature_of thy_name ::
930 term_features_of ctxt thy_name term_max_depth type_max_depth ts
931 |> scope <> Global ? cons local_feature
934 (* Too many dependencies is a sign that a decision procedure is at work. There is not much to learn
936 val max_dependencies = 20
938 val prover_default_max_facts = 25
940 (* "type_definition_xxx" facts are characterized by their use of "CollectI". *)
941 val typedef_dep = nickname_of_thm @{thm CollectI}
942 (* Mysterious parts of the class machinery create lots of proofs that refer exclusively to
943 "someI_ex" (and to some internal constructions). *)
944 val class_some_dep = nickname_of_thm @{thm someI_ex}
947 @{thms fundef_ex1_existence fundef_ex1_uniqueness fundef_ex1_iff fundef_default_value}
948 |> map nickname_of_thm
950 (* "Rep_xxx_inject", "Abs_xxx_inverse", etc., are derived using these facts. *)
952 @{thms type_definition.Abs_inverse type_definition.Rep_inverse type_definition.Rep
953 type_definition.Rep_inject type_definition.Abs_inject type_definition.Rep_cases
954 type_definition.Abs_cases type_definition.Rep_induct type_definition.Abs_induct
955 type_definition.Rep_range type_definition.Abs_image}
956 |> map nickname_of_thm
958 fun is_size_def [dep] th =
959 (case first_field ".rec" dep of
961 (case first_field ".size" (nickname_of_thm th) of
962 SOME (pref', _) => pref = pref'
965 | is_size_def _ _ = false
967 fun trim_dependencies deps =
968 if length deps > max_dependencies then NONE else SOME deps
970 fun isar_dependencies_of name_tabs th =
971 thms_in_proof max_dependencies (SOME name_tabs) th
972 |> Option.map (fn deps =>
973 if deps = [typedef_dep] orelse deps = [class_some_dep] orelse
974 exists (member (op =) fundef_ths) deps orelse exists (member (op =) typedef_ths) deps orelse
975 is_size_def deps th then
980 fun prover_dependencies_of ctxt (params as {verbose, max_facts, ...}) prover auto_level facts
982 (case isar_dependencies_of name_tabs th of
983 SOME [] => (false, [])
986 val isar_deps = these isar_deps0
987 val thy = Proof_Context.theory_of ctxt
988 val goal = goal_of_thm thy th
989 val name = nickname_of_thm th
990 val (_, hyp_ts, concl_t) = ATP_Util.strip_subgoal goal 1 ctxt
991 val facts = facts |> filter (fn (_, th') => thm_less (th', th))
993 fun nickify ((_, stature), th) = ((nickname_of_thm th, stature), th)
995 fun is_dep dep (_, th) = nickname_of_thm th = dep
997 fun add_isar_dep facts dep accum =
998 if exists (is_dep dep) accum then
1001 (case find_first (is_dep dep) facts of
1002 SOME ((_, status), th) => accum @ [(("", status), th)]
1003 | NONE => accum (* should not happen *))
1007 |> mepo_suggested_facts ctxt params (max_facts |> the_default prover_default_max_facts) NONE
1011 |> fold (add_isar_dep facts) isar_deps
1013 val num_isar_deps = length isar_deps
1015 if verbose andalso auto_level = 0 then
1016 writeln ("MaSh: " ^ quote prover ^ " on " ^ quote name ^ " with " ^
1017 string_of_int num_isar_deps ^ " + " ^ string_of_int (length facts - num_isar_deps) ^
1021 (case run_prover_for_mash ctxt params prover name facts goal of
1022 {outcome = NONE, used_facts, ...} =>
1023 (if verbose andalso auto_level = 0 then
1024 let val num_facts = length used_facts in
1025 writeln ("Found proof with " ^ string_of_int num_facts ^ " fact" ^
1026 plural_s num_facts ^ ".")
1030 (true, map fst used_facts))
1031 | _ => (false, isar_deps))
1035 (*** High-level communication with MaSh ***)
1037 (* In the following functions, chunks are risers w.r.t. "thm_less_eq". *)
1039 fun chunks_and_parents_for chunks th =
1041 fun insert_parent new parents =
1042 let val parents = parents |> filter_out (fn p => thm_less_eq (p, new)) in
1043 parents |> forall (fn p => not (thm_less_eq (new, p))) parents ? cons new
1046 fun rechunk seen (rest as th' :: ths) =
1047 if thm_less_eq (th', th) then (rev seen, rest)
1048 else rechunk (th' :: seen) ths
1050 fun do_chunk [] accum = accum
1051 | do_chunk (chunk as hd_chunk :: _) (chunks, parents) =
1052 if thm_less_eq (hd_chunk, th) then
1053 (chunk :: chunks, insert_parent hd_chunk parents)
1054 else if thm_less_eq (List.last chunk, th) then
1055 let val (front, back as hd_back :: _) = rechunk [] chunk in
1056 (front :: back :: chunks, insert_parent hd_back parents)
1059 (chunk :: chunks, parents)
1061 fold_rev do_chunk chunks ([], [])
1063 ||> map nickname_of_thm
1066 fun attach_parents_to_facts _ [] = []
1067 | attach_parents_to_facts old_facts (facts as (_, th) :: _) =
1069 fun do_facts _ [] = []
1070 | do_facts (_, parents) [fact] = [(parents, fact)]
1071 | do_facts (chunks, parents)
1072 ((fact as (_, th)) :: (facts as (_, th') :: _)) =
1074 val chunks = app_hd (cons th) chunks
1075 val chunks_and_parents' =
1076 if thm_less_eq (th, th') andalso thy_name_of_thm th = thy_name_of_thm th' then
1077 (chunks, [nickname_of_thm th])
1079 chunks_and_parents_for chunks th'
1081 (parents, fact) :: do_facts chunks_and_parents' facts
1085 |> do_facts (chunks_and_parents_for [[]] th)
1086 |> drop (length old_facts)
1089 fun maximal_wrt_graph G keys =
1091 val tab = Symtab.empty |> fold (fn name => Symtab.default (name, ())) keys
1093 fun insert_new seen name = not (Symtab.defined seen name) ? insert (op =) name
1095 fun num_keys keys = Graph.Keys.fold (K (Integer.add 1)) keys 0
1097 fun find_maxes _ (maxs, []) = map snd maxs
1098 | find_maxes seen (maxs, new :: news) =
1099 find_maxes (seen |> num_keys (Graph.imm_succs G new) > 1 ? Symtab.default (new, ()))
1100 (if Symtab.defined tab new then
1102 val newp = Graph.all_preds G [new]
1103 fun is_ancestor x yp = member (op =) yp x
1104 val maxs = maxs |> filter (fn (_, max) => not (is_ancestor max newp))
1106 if exists (is_ancestor new o fst) maxs then (maxs, news)
1107 else ((newp, new) :: filter_out (fn (_, max) => is_ancestor max newp) maxs, news)
1110 (maxs, Graph.Keys.fold (insert_new seen) (Graph.imm_preds G new) news))
1112 find_maxes Symtab.empty ([], Graph.maximals G)
1115 fun strict_subthy thyp = Theory.subthy thyp andalso not (Theory.subthy (swap thyp))
1117 fun maximal_wrt_access_graph _ [] = []
1118 | maximal_wrt_access_graph access_G ((fact as (_, th)) :: facts) =
1119 let val thy = Thm.theory_of_thm th in
1120 fact :: filter_out (fn (_, th') => strict_subthy (Thm.theory_of_thm th', thy)) facts
1121 |> map (nickname_of_thm o snd)
1122 |> maximal_wrt_graph access_G
1125 fun is_fact_in_graph access_G = can (Graph.get_node access_G) o nickname_of_thm
1127 val chained_feature_factor = 0.5 (* FUDGE *)
1128 val extra_feature_factor = 0.1 (* FUDGE *)
1129 val num_extra_feature_facts = 10 (* FUDGE *)
1131 val max_proximity_facts = 100 (* FUDGE *)
1133 fun find_mash_suggestions ctxt max_facts suggs facts chained raw_unknown =
1135 val inter_fact = inter (eq_snd Thm.eq_thm_prop)
1136 val raw_mash = find_suggested_facts ctxt facts suggs
1137 val proximate = take max_proximity_facts facts
1138 val unknown_chained = inter_fact raw_unknown chained
1139 val unknown_proximate = inter_fact raw_unknown proximate
1141 [(0.9 (* FUDGE *), (map (rpair 1.0) unknown_chained, [])),
1142 (0.4 (* FUDGE *), (weight_facts_smoothly unknown_proximate, [])),
1143 (0.1 (* FUDGE *), (weight_facts_steeply raw_mash, raw_unknown))]
1144 val unknown = raw_unknown
1145 |> fold (subtract (eq_snd Thm.eq_thm_prop)) [unknown_chained, unknown_proximate]
1147 (mesh_facts (eq_snd (gen_eq_thm ctxt)) max_facts mess, unknown)
1150 fun mash_suggested_facts ctxt thy ({debug, ...} : params) max_suggs hyp_ts concl_t facts =
1152 val thy_name = Context.theory_name thy
1153 val algorithm = the_mash_algorithm ()
1156 |> rev_sort_list_prefix (crude_thm_ord o apply2 snd)
1157 (Int.max (num_extra_feature_facts, max_proximity_facts))
1159 val chained = filter (fn ((_, (scope, _)), _) => scope = Chained) facts
1161 fun fact_has_right_theory (_, th) =
1162 thy_name = Context.theory_name (Thm.theory_of_thm th)
1164 fun chained_or_extra_features_of factor (((_, stature), th), weight) =
1166 |> features_of ctxt (Thm.theory_of_thm th) stature
1167 |> map (rpair (weight * factor))
1169 val {access_G, xtabs = ((num_facts, fact_tab), (num_feats, feat_tab)), ffds, freqs, ...} =
1172 val goal_feats0 = features_of ctxt thy (Local, General) (concl_t :: hyp_ts)
1173 val chained_feats = chained
1175 |> map (chained_or_extra_features_of chained_feature_factor)
1176 |> rpair [] |-> fold (union (eq_fst (op =)))
1177 val extra_feats = facts
1178 |> take (Int.max (0, num_extra_feature_facts - length chained))
1179 |> filter fact_has_right_theory
1180 |> weight_facts_steeply
1181 |> map (chained_or_extra_features_of extra_feature_factor)
1182 |> rpair [] |-> fold (union (eq_fst (op =)))
1185 fold (union (eq_fst (op =))) [chained_feats, extra_feats] (map (rpair 1.0) goal_feats0)
1186 |> debug ? sort (Real.compare o swap o apply2 snd)
1188 val parents = maximal_wrt_access_graph access_G facts
1189 val visible_facts = map_filter (Symtab.lookup fact_tab) (Graph.all_preds access_G parents)
1192 if algorithm = MaSh_NB_Ext orelse algorithm = MaSh_kNN_Ext then
1195 Graph.schedule (fn _ => fn (fact, (_, feats, deps)) => (fact, feats, deps)) access_G
1197 MaSh.query_external ctxt algorithm max_suggs learns goal_feats
1201 val int_goal_feats =
1202 map_filter (fn (s, w) => Option.map (rpair w) (Symtab.lookup feat_tab s)) goal_feats
1204 MaSh.query_internal ctxt algorithm num_facts num_feats ffds freqs visible_facts max_suggs
1205 goal_feats int_goal_feats
1208 val unknown = filter_out (is_fact_in_graph access_G o snd) facts
1210 find_mash_suggestions ctxt max_suggs suggs facts chained unknown
1211 |> apply2 (map fact_of_raw_fact)
1214 fun mash_unlearn () = (clear_state (); writeln "Reset MaSh.")
1216 fun learn_wrt_access_graph ctxt (name, parents, feats, deps)
1217 (accum as (access_G, (fact_xtab, feat_xtab))) =
1219 fun maybe_learn_from from (accum as (parents, access_G)) =
1220 try_graph ctxt "updating graph" accum (fn () =>
1221 (from :: parents, Graph.add_edge_acyclic (from, name) access_G))
1223 val access_G = access_G |> Graph.default_node (name, (Isar_Proof, feats, deps))
1224 val (parents, access_G) = ([], access_G) |> fold maybe_learn_from parents
1225 val (deps, _) = ([], access_G) |> fold maybe_learn_from deps
1227 val fact_xtab = add_to_xtab name fact_xtab
1228 val feat_xtab = fold maybe_add_to_xtab feats feat_xtab
1230 (SOME (name, parents, feats, deps), (access_G, (fact_xtab, feat_xtab)))
1232 handle Symtab.DUP _ => (NONE, accum) (* facts sometimes have the same name, confusingly *)
1234 fun relearn_wrt_access_graph ctxt (name, deps) access_G =
1236 fun maybe_relearn_from from (accum as (parents, access_G)) =
1237 try_graph ctxt "updating graph" accum (fn () =>
1238 (from :: parents, Graph.add_edge_acyclic (from, name) access_G))
1240 access_G |> Graph.map_node name (fn (_, feats, _) => (Automatic_Proof, feats, deps))
1241 val (deps, _) = ([], access_G) |> fold maybe_relearn_from deps
1243 ((name, deps), access_G)
1246 fun flop_wrt_access_graph name =
1247 Graph.map_node name (fn (_, feats, deps) => (Isar_Proof_wegen_Prover_Flop, feats, deps))
1249 val learn_timeout_slack = 20.0
1251 fun launch_thread timeout task =
1253 val hard_timeout = time_mult learn_timeout_slack timeout
1254 val birth_time = Time.now ()
1255 val death_time = Time.+ (birth_time, hard_timeout)
1256 val desc = ("Machine learner for Sledgehammer", "")
1258 Async_Manager_Legacy.thread MaShN birth_time death_time desc task
1261 fun learned_proof_name () =
1262 Date.fmt ".%Y%m%d.%H%M%S." (Date.fromTimeLocal (Time.now ())) ^ serial_string ()
1264 fun mash_learn_proof ctxt ({timeout, ...} : params) t facts used_ths =
1265 if not (null used_ths) andalso is_mash_enabled () then
1266 launch_thread timeout (fn () =>
1268 val thy = Proof_Context.theory_of ctxt
1269 val feats = features_of ctxt thy (Local, General) [t]
1270 val facts = rev_sort_list_prefix (crude_thm_ord o apply2 snd) 1 facts
1273 (fn {access_G, xtabs as ((num_facts0, _), _), ffds, freqs, dirty_facts} =>
1275 val parents = maximal_wrt_access_graph access_G facts
1277 |> filter (is_fact_in_graph access_G)
1278 |> map nickname_of_thm
1280 val name = learned_proof_name ()
1281 val (access_G', xtabs', rev_learns) =
1282 add_node Automatic_Proof name parents feats deps (access_G, xtabs, [])
1284 val (ffds', freqs') =
1285 recompute_ffds_freqs_from_learns (rev rev_learns) xtabs' num_facts0 ffds freqs
1287 {access_G = access_G', xtabs = xtabs', ffds = ffds', freqs = freqs',
1288 dirty_facts = Option.map (cons name) dirty_facts}
1295 fun sendback sub = Active.sendback_markup [Markup.padding_command] (sledgehammerN ^ " " ^ sub)
1297 val commit_timeout = seconds 30.0
1299 (* The timeout is understood in a very relaxed fashion. *)
1300 fun mash_learn_facts ctxt (params as {debug, verbose, ...}) prover auto_level run_prover
1301 learn_timeout facts =
1303 val timer = Timer.startRealTimer ()
1304 fun next_commit_time () = Time.+ (Timer.checkRealTimer timer, commit_timeout)
1306 val {access_G, ...} = peek_state ctxt
1307 val is_in_access_G = is_fact_in_graph access_G o snd
1308 val no_new_facts = forall is_in_access_G facts
1310 if no_new_facts andalso not run_prover then
1311 if auto_level < 2 then
1312 "No new " ^ (if run_prover then "automatic" else "Isar") ^ " proofs to learn." ^
1313 (if auto_level = 0 andalso not run_prover then
1314 "\n\nHint: Try " ^ sendback learn_proverN ^ " to learn from an automatic prover."
1321 val name_tabs = build_name_tables nickname_of_thm facts
1323 fun deps_of status th =
1324 if status = Non_Rec_Def orelse status = Rec_Def then
1326 else if run_prover then
1327 prover_dependencies_of ctxt params prover auto_level facts name_tabs th
1328 |> (fn (false, _) => NONE | (true, deps) => trim_dependencies deps)
1330 isar_dependencies_of name_tabs th
1332 fun do_commit [] [] [] state = state
1333 | do_commit learns relearns flops
1334 {access_G, xtabs as ((num_facts0, _), _), ffds, freqs, dirty_facts} =
1336 val was_empty = Graph.is_empty access_G
1338 val (learns, (access_G', xtabs')) =
1339 fold_map (learn_wrt_access_graph ctxt) learns (access_G, xtabs)
1341 val (relearns, access_G'') =
1342 fold_map (relearn_wrt_access_graph ctxt) relearns access_G'
1344 val access_G''' = access_G'' |> fold flop_wrt_access_graph flops
1346 (case (was_empty, dirty_facts) of
1347 (false, SOME names) => SOME (map #1 learns @ map #1 relearns @ names)
1350 val (ffds', freqs') =
1351 if null relearns then
1352 recompute_ffds_freqs_from_learns
1353 (map (fn (name, _, feats, deps) => (name, feats, deps)) learns) xtabs' num_facts0
1356 recompute_ffds_freqs_from_access_G access_G''' xtabs'
1358 {access_G = access_G''', xtabs = xtabs', ffds = ffds', freqs = freqs',
1359 dirty_facts = dirty_facts'}
1362 fun commit last learns relearns flops =
1363 (if debug andalso auto_level = 0 then writeln "Committing..." else ();
1364 map_state ctxt (do_commit (rev learns) relearns flops);
1365 if not last andalso auto_level = 0 then
1366 let val num_proofs = length learns + length relearns in
1367 writeln ("Learned " ^ string_of_int num_proofs ^ " " ^
1368 (if run_prover then "automatic" else "Isar") ^ " proof" ^
1369 plural_s num_proofs ^ " in the last " ^ string_of_time commit_timeout ^ ".")
1374 fun learn_new_fact _ (accum as (_, (_, _, true))) = accum
1375 | learn_new_fact (parents, ((_, stature as (_, status)), th))
1376 (learns, (num_nontrivial, next_commit, _)) =
1378 val name = nickname_of_thm th
1379 val feats = features_of ctxt (Thm.theory_of_thm th) stature [Thm.prop_of th]
1380 val deps = these (deps_of status th)
1381 val num_nontrivial = num_nontrivial |> not (null deps) ? Integer.add 1
1382 val learns = (name, parents, feats, deps) :: learns
1383 val (learns, next_commit) =
1384 if Time.> (Timer.checkRealTimer timer, next_commit) then
1385 (commit false learns [] []; ([], next_commit_time ()))
1387 (learns, next_commit)
1388 val timed_out = Time.> (Timer.checkRealTimer timer, learn_timeout)
1390 (learns, (num_nontrivial, next_commit, timed_out))
1393 val (num_new_facts, num_nontrivial) =
1394 if no_new_facts then
1398 val new_facts = facts
1399 |> sort (crude_thm_ord o apply2 snd)
1400 |> attach_parents_to_facts []
1401 |> filter_out (is_in_access_G o snd)
1402 val (learns, (num_nontrivial, _, _)) =
1403 ([], (0, next_commit_time (), false))
1404 |> fold learn_new_fact new_facts
1406 commit true learns [] []; (length new_facts, num_nontrivial)
1409 fun relearn_old_fact _ (accum as (_, (_, _, true))) = accum
1410 | relearn_old_fact ((_, (_, status)), th)
1411 ((relearns, flops), (num_nontrivial, next_commit, _)) =
1413 val name = nickname_of_thm th
1414 val (num_nontrivial, relearns, flops) =
1415 (case deps_of status th of
1416 SOME deps => (num_nontrivial + 1, (name, deps) :: relearns, flops)
1417 | NONE => (num_nontrivial, relearns, name :: flops))
1418 val (relearns, flops, next_commit) =
1419 if Time.> (Timer.checkRealTimer timer, next_commit) then
1420 (commit false [] relearns flops; ([], [], next_commit_time ()))
1422 (relearns, flops, next_commit)
1423 val timed_out = Time.> (Timer.checkRealTimer timer, learn_timeout)
1425 ((relearns, flops), (num_nontrivial, next_commit, timed_out))
1428 val num_nontrivial =
1429 if not run_prover then
1433 val max_isar = 1000 * max_dependencies
1435 fun priority_of th =
1436 Random.random_range 0 max_isar +
1437 (case try (Graph.get_node access_G) (nickname_of_thm th) of
1438 SOME (Isar_Proof, _, deps) => ~100 * length deps
1439 | SOME (Automatic_Proof, _, _) => 2 * max_isar
1440 | SOME (Isar_Proof_wegen_Prover_Flop, _, _) => max_isar
1443 val old_facts = facts
1444 |> filter is_in_access_G
1445 |> map (`(priority_of o snd))
1446 |> sort (int_ord o apply2 fst)
1448 val ((relearns, flops), (num_nontrivial, _, _)) =
1449 (([], []), (num_nontrivial, next_commit_time (), false))
1450 |> fold relearn_old_fact old_facts
1452 commit true [] relearns flops; num_nontrivial
1455 if verbose orelse auto_level < 2 then
1456 "Learned " ^ string_of_int num_new_facts ^ " fact" ^ plural_s num_new_facts ^ " and " ^
1457 string_of_int num_nontrivial ^ " nontrivial " ^
1458 (if run_prover then "automatic and " else "") ^ "Isar proof" ^ plural_s num_nontrivial ^
1459 (if verbose then " in " ^ string_of_time (Timer.checkRealTimer timer) else "") ^ "."
1465 fun mash_learn ctxt (params as {provers, timeout, ...}) fact_override chained run_prover =
1467 val css = Sledgehammer_Fact.clasimpset_rule_table_of ctxt
1468 val ctxt = ctxt |> Config.put instantiate_inducts false
1470 nearly_all_facts ctxt false fact_override Keyword.empty_keywords css chained [] @{prop True}
1471 |> sort (crude_thm_ord o apply2 snd o swap)
1472 val num_facts = length facts
1473 val prover = hd provers
1475 fun learn auto_level run_prover =
1476 mash_learn_facts ctxt params prover auto_level run_prover one_year facts
1480 (writeln ("MaShing through " ^ string_of_int num_facts ^ " fact" ^
1481 plural_s num_facts ^ " for automatic proofs (" ^ quote prover ^ " timeout: " ^
1482 string_of_time timeout ^ ").\n\nCollecting Isar proofs first...");
1484 writeln "Now collecting automatic proofs. This may take several hours. You \
1485 \can safely stop the learning process at any point.";
1488 (writeln ("MaShing through " ^ string_of_int num_facts ^ " fact" ^
1489 plural_s num_facts ^ " for Isar proofs...");
1493 fun mash_can_suggest_facts ctxt = not (Graph.is_empty (#access_G (peek_state ctxt)))
1495 (* Generate more suggestions than requested, because some might be thrown out later for various
1496 reasons (e.g., duplicates). *)
1497 fun generous_max_suggestions max_facts = 3 * max_facts div 2 + 25
1499 val mepo_weight = 0.5
1500 val mash_weight = 0.5
1502 val max_facts_to_learn_before_query = 100
1504 (* The threshold should be large enough so that MaSh does not get activated for Auto Sledgehammer
1506 val min_secs_for_learning = 15
1508 fun relevant_facts ctxt (params as {verbose, learn, fact_filter, timeout, ...}) prover
1509 max_facts ({add, only, ...} : fact_override) hyp_ts concl_t facts =
1510 if not (subset (op =) (the_list fact_filter, fact_filters)) then
1511 error ("Unknown fact filter: " ^ quote (the fact_filter) ^ ".")
1513 [("", map fact_of_raw_fact facts)]
1514 else if max_facts <= 0 orelse null facts then
1518 val thy = Proof_Context.theory_of ctxt
1520 fun maybe_launch_thread exact min_num_facts_to_learn =
1521 if not (Async_Manager_Legacy.has_running_threads MaShN) andalso
1522 Time.toSeconds timeout >= min_secs_for_learning then
1523 let val timeout = time_mult learn_timeout_slack timeout in
1525 writeln ("Started MaShing through " ^
1526 (if exact then "" else "at least ") ^ string_of_int min_num_facts_to_learn ^
1527 " fact" ^ plural_s min_num_facts_to_learn ^ " in the background.")
1530 launch_thread timeout
1531 (fn () => (true, mash_learn_facts ctxt params prover 2 false timeout facts))
1536 fun please_learn () =
1538 val {access_G, xtabs = ((num_facts0, _), _), ...} = peek_state ctxt
1539 val is_in_access_G = is_fact_in_graph access_G o snd
1540 val min_num_facts_to_learn = length facts - num_facts0
1542 if min_num_facts_to_learn <= max_facts_to_learn_before_query then
1543 (case length (filter_out is_in_access_G facts) of
1545 | num_facts_to_learn =>
1546 if num_facts_to_learn <= max_facts_to_learn_before_query then
1547 mash_learn_facts ctxt params prover 2 false timeout facts
1548 |> (fn "" => () | s => writeln (MaShN ^ ": " ^ s))
1550 maybe_launch_thread true num_facts_to_learn)
1552 maybe_launch_thread false min_num_facts_to_learn
1555 val mash_enabled = is_mash_enabled ()
1557 if learn andalso mash_enabled andalso fact_filter <> SOME mepoN then please_learn () else ()
1559 val effective_fact_filter =
1560 (case fact_filter of
1562 | NONE => if mash_enabled andalso mash_can_suggest_facts ctxt then meshN else mepoN)
1564 val unique_facts = drop_duplicate_facts facts
1565 val add_ths = Attrib.eval_thms ctxt add
1567 fun in_add (_, th) = member Thm.eq_thm_prop add_ths th
1569 fun add_and_take accepts =
1573 (unique_facts |> filter in_add |> map fact_of_raw_fact) @ (accepts |> filter_out in_add))
1577 (mepo_suggested_facts ctxt params max_facts NONE hyp_ts concl_t unique_facts
1578 |> weight_facts_steeply, [])
1581 mash_suggested_facts ctxt thy params (generous_max_suggestions max_facts) hyp_ts concl_t
1583 |>> weight_facts_steeply
1586 (* the order is important for the "case" expression below *)
1587 [] |> effective_fact_filter <> mepoN ? cons (mash_weight, mash)
1588 |> effective_fact_filter <> mashN ? cons (mepo_weight, mepo)
1589 |> Par_List.map (apsnd (fn f => f ()))
1590 val mesh = mesh_facts (eq_snd (gen_eq_thm ctxt)) max_facts mess |> add_and_take
1592 (case (fact_filter, mess) of
1593 (NONE, [(_, (mepo, _)), (_, (mash, _))]) =>
1595 (mepoN, mepo |> map fst |> add_and_take),
1596 (mashN, mash |> map fst |> add_and_take)]
1597 | _ => [(effective_fact_filter, mesh)])
1600 fun kill_learners () = Async_Manager_Legacy.kill_threads MaShN "learner"
1601 fun running_learners () = Async_Manager_Legacy.running_threads MaShN "learner"