src/HOL/Tools/Sledgehammer/sledgehammer_filter_mash.ML
author blanchet
Fri, 20 Jul 2012 22:19:45 +0200
changeset 49393 9e96486d53ad
parent 49392 4a11914fd530
child 49394 2b5ad61e2ccc
permissions -rw-r--r--
handle local facts smoothly in MaSh
blanchet@49263
     1
(*  Title:      HOL/Tools/Sledgehammer/sledgehammer_filter_mash.ML
blanchet@49263
     2
    Author:     Jasmin Blanchette, TU Muenchen
blanchet@49263
     3
blanchet@49263
     4
Sledgehammer's machine-learning-based relevance filter (MaSh).
blanchet@49263
     5
*)
blanchet@49263
     6
blanchet@49263
     7
signature SLEDGEHAMMER_FILTER_MASH =
blanchet@49263
     8
sig
blanchet@49266
     9
  type status = ATP_Problem_Generate.status
blanchet@49266
    10
  type stature = ATP_Problem_Generate.stature
blanchet@49311
    11
  type fact = Sledgehammer_Fact.fact
blanchet@49311
    12
  type fact_override = Sledgehammer_Fact.fact_override
blanchet@49266
    13
  type params = Sledgehammer_Provers.params
blanchet@49303
    14
  type relevance_fudge = Sledgehammer_Provers.relevance_fudge
blanchet@49266
    15
  type prover_result = Sledgehammer_Provers.prover_result
blanchet@49266
    16
blanchet@49323
    17
  val trace : bool Config.T
blanchet@49334
    18
  val MaShN : string
blanchet@49329
    19
  val meshN : string
blanchet@49329
    20
  val iterN : string
blanchet@49329
    21
  val mashN : string
blanchet@49329
    22
  val fact_filters : string list
blanchet@49318
    23
  val escape_meta : string -> string
blanchet@49318
    24
  val escape_metas : string list -> string
blanchet@49323
    25
  val unescape_meta : string -> string
blanchet@49323
    26
  val unescape_metas : string -> string list
blanchet@49326
    27
  val extract_query : string -> string * string list
blanchet@49393
    28
  val nickname_of : thm -> string
blanchet@49336
    29
  val suggested_facts : string list -> ('a * thm) list -> ('a * thm) list
blanchet@49336
    30
  val mesh_facts :
blanchet@49336
    31
    int -> (('a * thm) list * ('a * thm) list) list -> ('a * thm) list
blanchet@49339
    32
  val is_likely_tautology_or_too_meta : thm -> bool
blanchet@49266
    33
  val theory_ord : theory * theory -> order
blanchet@49266
    34
  val thm_ord : thm * thm -> order
blanchet@49333
    35
  val features_of :
blanchet@49333
    36
    Proof.context -> string -> theory -> status -> term list -> string list
blanchet@49331
    37
  val isabelle_dependencies_of : unit Symtab.table -> thm -> string list
blanchet@49266
    38
  val goal_of_thm : theory -> thm -> thm
blanchet@49336
    39
  val run_prover_for_mash :
blanchet@49333
    40
    Proof.context -> params -> string -> fact list -> thm -> prover_result
blanchet@49347
    41
  val mash_CLEAR : Proof.context -> unit
blanchet@49331
    42
  val mash_INIT :
blanchet@49331
    43
    Proof.context -> bool
blanchet@49331
    44
    -> (string * string list * string list * string list) list -> unit
blanchet@49323
    45
  val mash_ADD :
blanchet@49331
    46
    Proof.context -> bool
blanchet@49331
    47
    -> (string * string list * string list * string list) list -> unit
blanchet@49331
    48
  val mash_QUERY :
blanchet@49333
    49
    Proof.context -> bool -> int -> string list * string list -> string list
blanchet@49347
    50
  val mash_unlearn : Proof.context -> unit
blanchet@49333
    51
  val mash_could_suggest_facts : unit -> bool
blanchet@49336
    52
  val mash_can_suggest_facts : Proof.context -> bool
blanchet@49313
    53
  val mash_suggest_facts :
blanchet@49336
    54
    Proof.context -> params -> string -> int -> term list -> term
blanchet@49336
    55
    -> ('a * thm) list -> ('a * thm) list * ('a * thm) list
blanchet@49334
    56
  val mash_learn_thy :
blanchet@49334
    57
    Proof.context -> params -> theory -> Time.time -> fact list -> string
blanchet@49331
    58
  val mash_learn_proof :
blanchet@49336
    59
    Proof.context -> params -> term -> ('a * thm) list -> thm list -> unit
blanchet@49303
    60
  val relevant_facts :
blanchet@49307
    61
    Proof.context -> params -> string -> int -> fact_override -> term list
blanchet@49311
    62
    -> term -> fact list -> fact list
blanchet@49334
    63
  val kill_learners : unit -> unit
blanchet@49334
    64
  val running_learners : unit -> unit
blanchet@49263
    65
end;
blanchet@49263
    66
blanchet@49263
    67
structure Sledgehammer_Filter_MaSh : SLEDGEHAMMER_FILTER_MASH =
blanchet@49263
    68
struct
blanchet@49264
    69
blanchet@49266
    70
open ATP_Util
blanchet@49266
    71
open ATP_Problem_Generate
blanchet@49266
    72
open Sledgehammer_Util
blanchet@49266
    73
open Sledgehammer_Fact
blanchet@49266
    74
open Sledgehammer_Filter_Iter
blanchet@49266
    75
open Sledgehammer_Provers
blanchet@49333
    76
open Sledgehammer_Minimize
blanchet@49266
    77
blanchet@49323
    78
val trace =
blanchet@49323
    79
  Attrib.setup_config_bool @{binding sledgehammer_filter_mash_trace} (K false)
blanchet@49323
    80
fun trace_msg ctxt msg = if Config.get ctxt trace then tracing (msg ()) else ()
blanchet@49323
    81
blanchet@49334
    82
val MaShN = "MaSh"
blanchet@49334
    83
blanchet@49329
    84
val meshN = "mesh"
blanchet@49329
    85
val iterN = "iter"
blanchet@49329
    86
val mashN = "mash"
blanchet@49329
    87
blanchet@49329
    88
val fact_filters = [meshN, iterN, mashN]
blanchet@49329
    89
blanchet@49329
    90
fun mash_home () = getenv "MASH_HOME"
blanchet@49329
    91
fun mash_state_dir () =
blanchet@49324
    92
  getenv "ISABELLE_HOME_USER" ^ "/mash"
blanchet@49331
    93
  |> tap (Isabelle_System.mkdir o Path.explode)
blanchet@49329
    94
fun mash_state_path () = mash_state_dir () ^ "/state" |> Path.explode
blanchet@49266
    95
blanchet@49345
    96
blanchet@49266
    97
(*** Isabelle helpers ***)
blanchet@49266
    98
blanchet@49323
    99
fun meta_char c =
blanchet@49266
   100
  if Char.isAlphaNum c orelse c = #"_" orelse c = #"." orelse c = #"(" orelse
blanchet@49266
   101
     c = #")" orelse c = #"," then
blanchet@49266
   102
    String.str c
blanchet@49266
   103
  else
blanchet@49266
   104
    (* fixed width, in case more digits follow *)
blanchet@49266
   105
    "\\" ^ stringN_of_int 3 (Char.ord c)
blanchet@49266
   106
blanchet@49323
   107
fun unmeta_chars accum [] = String.implode (rev accum)
blanchet@49323
   108
  | unmeta_chars accum (#"\\" :: d1 :: d2 :: d3 :: cs) =
blanchet@49323
   109
    (case Int.fromString (String.implode [d1, d2, d3]) of
blanchet@49323
   110
       SOME n => unmeta_chars (Char.chr n :: accum) cs
blanchet@49323
   111
     | NONE => "" (* error *))
blanchet@49323
   112
  | unmeta_chars _ (#"\\" :: _) = "" (* error *)
blanchet@49323
   113
  | unmeta_chars accum (c :: cs) = unmeta_chars (c :: accum) cs
blanchet@49323
   114
blanchet@49323
   115
val escape_meta = String.translate meta_char
blanchet@49318
   116
val escape_metas = map escape_meta #> space_implode " "
blanchet@49330
   117
val unescape_meta = String.explode #> unmeta_chars []
blanchet@49330
   118
val unescape_metas =
blanchet@49330
   119
  space_explode " " #> filter_out (curry (op =) "") #> map unescape_meta
blanchet@49266
   120
blanchet@49326
   121
fun extract_query line =
blanchet@49326
   122
  case space_explode ":" line of
blanchet@49330
   123
    [goal_name, suggs] => (unescape_meta goal_name, unescape_metas suggs)
blanchet@49327
   124
  | _ => ("", [])
blanchet@49326
   125
blanchet@49393
   126
fun parent_of_local_thm th =
blanchet@49393
   127
  let
blanchet@49393
   128
    val thy = th |> Thm.theory_of_thm
blanchet@49393
   129
    val facts = thy |> Global_Theory.facts_of
blanchet@49393
   130
    val space = facts |> Facts.space_of
blanchet@49393
   131
    fun id_of s = #id (Name_Space.the_entry space s)
blanchet@49393
   132
    fun max_id (s', _) (s, id) =
blanchet@49393
   133
      let val id' = id_of s' in if id > id' then (s, id) else (s', id') end
blanchet@49393
   134
  in ("", ~1) |> Facts.fold_static max_id facts |> fst end
blanchet@49393
   135
blanchet@49393
   136
val local_prefix = "local" ^ Long_Name.separator
blanchet@49393
   137
blanchet@49393
   138
fun nickname_of th =
blanchet@49393
   139
  let val hint = Thm.get_name_hint th in
blanchet@49393
   140
    (* FIXME: There must be a better way to detect local facts. *)
blanchet@49393
   141
    case try (unprefix local_prefix) hint of
blanchet@49393
   142
      SOME suff =>
blanchet@49393
   143
      parent_of_local_thm th ^ Long_Name.separator ^ Long_Name.separator ^ suff
blanchet@49393
   144
    | NONE => hint
blanchet@49393
   145
  end
blanchet@49393
   146
blanchet@49345
   147
fun suggested_facts suggs facts =
blanchet@49345
   148
  let
blanchet@49393
   149
    fun add_fact (fact as (_, th)) = Symtab.default (nickname_of th, fact)
blanchet@49345
   150
    val tab = Symtab.empty |> fold add_fact facts
blanchet@49345
   151
  in map_filter (Symtab.lookup tab) suggs end
blanchet@49326
   152
blanchet@49344
   153
(* Ad hoc score function roughly based on Blanchette's Ringberg 2011 data. *)
blanchet@49344
   154
fun score x = Math.pow (1.5, 15.5 - 0.05 * Real.fromInt x) + 15.0
blanchet@49343
   155
blanchet@49343
   156
fun sum_sq_avg [] = 0
blanchet@49344
   157
  | sum_sq_avg xs =
blanchet@49344
   158
    Real.ceil (100000.0 * fold (curry (op +) o score) xs 0.0) div length xs
blanchet@49328
   159
blanchet@49335
   160
fun mesh_facts max_facts [(selected, unknown)] =
blanchet@49335
   161
    take max_facts selected @ take (max_facts - length selected) unknown
blanchet@49329
   162
  | mesh_facts max_facts mess =
blanchet@49329
   163
    let
blanchet@49335
   164
      val mess = mess |> map (apfst (`length))
blanchet@49329
   165
      val fact_eq = Thm.eq_thm o pairself snd
blanchet@49335
   166
      fun score_in fact ((sel_len, sels), unks) =
blanchet@49335
   167
        case find_index (curry fact_eq fact) sels of
blanchet@49335
   168
          ~1 => (case find_index (curry fact_eq fact) unks of
blanchet@49344
   169
                   ~1 => SOME sel_len
blanchet@49335
   170
                 | _ => NONE)
blanchet@49344
   171
        | j => SOME j
blanchet@49343
   172
      fun score_of fact = mess |> map_filter (score_in fact) |> sum_sq_avg
blanchet@49335
   173
      val facts = fold (union fact_eq o take max_facts o snd o fst) mess []
blanchet@49329
   174
    in
blanchet@49343
   175
      facts |> map (`score_of) |> sort (int_ord o swap o pairself fst)
blanchet@49343
   176
            |> map snd |> take max_facts
blanchet@49329
   177
    end
blanchet@49327
   178
blanchet@49318
   179
val thy_feature_prefix = "y_"
blanchet@49266
   180
blanchet@49318
   181
val thy_feature_name_of = prefix thy_feature_prefix
blanchet@49318
   182
val const_name_of = prefix const_prefix
blanchet@49318
   183
val type_name_of = prefix type_const_prefix
blanchet@49318
   184
val class_name_of = prefix class_prefix
blanchet@49266
   185
blanchet@49339
   186
fun is_likely_tautology_or_too_meta th =
blanchet@49339
   187
  let
blanchet@49339
   188
    val is_boring_const = member (op =) atp_widely_irrelevant_consts
blanchet@49339
   189
    fun is_boring_bool t =
blanchet@49339
   190
      not (exists_Const (not o is_boring_const o fst) t) orelse
blanchet@49339
   191
      exists_type (exists_subtype (curry (op =) @{typ prop})) t
blanchet@49339
   192
    fun is_boring_prop (@{const Trueprop} $ t) = is_boring_bool t
blanchet@49339
   193
      | is_boring_prop (@{const "==>"} $ t $ u) =
blanchet@49339
   194
        is_boring_prop t andalso is_boring_prop u
blanchet@49339
   195
      | is_boring_prop (Const (@{const_name all}, _) $ (Abs (_, _, t)) $ u) =
blanchet@49339
   196
        is_boring_prop t andalso is_boring_prop u
blanchet@49339
   197
      | is_boring_prop (Const (@{const_name "=="}, _) $ t $ u) =
blanchet@49339
   198
        is_boring_bool t andalso is_boring_bool u
blanchet@49339
   199
      | is_boring_prop _ = true
blanchet@49339
   200
  in
blanchet@49339
   201
    is_boring_prop (prop_of th) andalso not (Thm.eq_thm_prop (@{thm ext}, th))
blanchet@49339
   202
  end
blanchet@49339
   203
blanchet@49339
   204
fun theory_ord p =
blanchet@49339
   205
  if Theory.eq_thy p then
blanchet@49339
   206
    EQUAL
blanchet@49339
   207
  else if Theory.subthy p then
blanchet@49339
   208
    LESS
blanchet@49339
   209
  else if Theory.subthy (swap p) then
blanchet@49339
   210
    GREATER
blanchet@49339
   211
  else case int_ord (pairself (length o Theory.ancestors_of) p) of
blanchet@49339
   212
    EQUAL => string_ord (pairself Context.theory_name p)
blanchet@49339
   213
  | order => order
blanchet@49339
   214
blanchet@49339
   215
val thm_ord = theory_ord o pairself theory_of_thm
blanchet@49339
   216
blanchet@49341
   217
val bad_types = [@{type_name prop}, @{type_name bool}, @{type_name fun}]
blanchet@49341
   218
blanchet@49333
   219
fun interesting_terms_types_and_classes ctxt prover term_max_depth
blanchet@49333
   220
                                        type_max_depth ts =
blanchet@49266
   221
  let
blanchet@49333
   222
    fun is_bad_const (x as (s, _)) args =
blanchet@49333
   223
      member (op =) atp_logical_consts s orelse
blanchet@49333
   224
      fst (is_built_in_const_for_prover ctxt prover x args)
blanchet@49319
   225
    fun add_classes @{sort type} = I
blanchet@49319
   226
      | add_classes S = union (op =) (map class_name_of S)
blanchet@49266
   227
    fun do_add_type (Type (s, Ts)) =
blanchet@49266
   228
        (not (member (op =) bad_types s) ? insert (op =) (type_name_of s))
blanchet@49266
   229
        #> fold do_add_type Ts
blanchet@49319
   230
      | do_add_type (TFree (_, S)) = add_classes S
blanchet@49319
   231
      | do_add_type (TVar (_, S)) = add_classes S
blanchet@49266
   232
    fun add_type T = type_max_depth >= 0 ? do_add_type T
blanchet@49266
   233
    fun mk_app s args =
blanchet@49266
   234
      if member (op <>) args "" then s ^ "(" ^ space_implode "," args ^ ")"
blanchet@49266
   235
      else s
blanchet@49266
   236
    fun patternify ~1 _ = ""
blanchet@49266
   237
      | patternify depth t =
blanchet@49266
   238
        case strip_comb t of
blanchet@49266
   239
          (Const (s, _), args) =>
blanchet@49266
   240
          mk_app (const_name_of s) (map (patternify (depth - 1)) args)
blanchet@49266
   241
        | _ => ""
blanchet@49266
   242
    fun add_term_patterns ~1 _ = I
blanchet@49266
   243
      | add_term_patterns depth t =
blanchet@49266
   244
        insert (op =) (patternify depth t)
blanchet@49266
   245
        #> add_term_patterns (depth - 1) t
blanchet@49266
   246
    val add_term = add_term_patterns term_max_depth
blanchet@49266
   247
    fun add_patterns t =
blanchet@49266
   248
      let val (head, args) = strip_comb t in
blanchet@49266
   249
        (case head of
blanchet@49333
   250
           Const (x as (_, T)) =>
blanchet@49333
   251
           not (is_bad_const x args) ? (add_term t #> add_type T)
blanchet@49266
   252
         | Free (_, T) => add_type T
blanchet@49266
   253
         | Var (_, T) => add_type T
blanchet@49266
   254
         | Abs (_, T, body) => add_type T #> add_patterns body
blanchet@49266
   255
         | _ => I)
blanchet@49266
   256
        #> fold add_patterns args
blanchet@49266
   257
      end
blanchet@49341
   258
  in [] |> fold add_patterns ts end
blanchet@49266
   259
blanchet@49266
   260
fun is_exists (s, _) = (s = @{const_name Ex} orelse s = @{const_name Ex1})
blanchet@49266
   261
blanchet@49312
   262
val term_max_depth = 1
blanchet@49312
   263
val type_max_depth = 1
blanchet@49266
   264
blanchet@49266
   265
(* TODO: Generate type classes for types? *)
blanchet@49333
   266
fun features_of ctxt prover thy status ts =
blanchet@49318
   267
  thy_feature_name_of (Context.theory_name thy) ::
blanchet@49333
   268
  interesting_terms_types_and_classes ctxt prover term_max_depth type_max_depth
blanchet@49333
   269
                                      ts
blanchet@49347
   270
  |> forall is_lambda_free ts ? cons "no_lams"
blanchet@49347
   271
  |> forall (not o exists_Const is_exists) ts ? cons "no_skos"
blanchet@49317
   272
  |> (case status of
blanchet@49317
   273
        General => I
blanchet@49317
   274
      | Induction => cons "induction"
blanchet@49317
   275
      | Intro => cons "intro"
blanchet@49317
   276
      | Inductive => cons "inductive"
blanchet@49317
   277
      | Elim => cons "elim"
blanchet@49317
   278
      | Simp => cons "simp"
blanchet@49317
   279
      | Def => cons "def")
blanchet@49266
   280
blanchet@49341
   281
fun isabelle_dependencies_of all_facts = thms_in_proof (SOME all_facts)
blanchet@49266
   282
blanchet@49266
   283
val freezeT = Type.legacy_freeze_type
blanchet@49266
   284
blanchet@49266
   285
fun freeze (t $ u) = freeze t $ freeze u
blanchet@49266
   286
  | freeze (Abs (s, T, t)) = Abs (s, freezeT T, freeze t)
blanchet@49266
   287
  | freeze (Var ((s, _), T)) = Free (s, freezeT T)
blanchet@49266
   288
  | freeze (Const (s, T)) = Const (s, freezeT T)
blanchet@49266
   289
  | freeze (Free (s, T)) = Free (s, freezeT T)
blanchet@49266
   290
  | freeze t = t
blanchet@49266
   291
blanchet@49266
   292
fun goal_of_thm thy = prop_of #> freeze #> cterm_of thy #> Goal.init
blanchet@49266
   293
blanchet@49336
   294
fun run_prover_for_mash ctxt params prover facts goal =
blanchet@49266
   295
  let
blanchet@49266
   296
    val problem =
blanchet@49266
   297
      {state = Proof.init ctxt, goal = goal, subgoal = 1, subgoal_count = 1,
blanchet@49304
   298
       facts = facts |> map (apfst (apfst (fn name => name ())))
blanchet@49333
   299
                     |> map Untranslated_Fact}
blanchet@49336
   300
    val prover = get_minimizing_prover ctxt Normal (K ()) prover
blanchet@49266
   301
  in prover params (K (K (K ""))) problem end
blanchet@49266
   302
blanchet@49266
   303
blanchet@49317
   304
(*** Low-level communication with MaSh ***)
blanchet@49317
   305
blanchet@49338
   306
fun write_file (xs, f) file =
blanchet@49333
   307
  let val path = Path.explode file in
blanchet@49338
   308
    File.write path "";
blanchet@49338
   309
    xs |> chunk_list 500
blanchet@49338
   310
       |> List.app (File.append path o space_implode "" o map f)
blanchet@49333
   311
  end
blanchet@49331
   312
blanchet@49331
   313
fun mash_info overlord =
blanchet@49331
   314
  if overlord then (getenv "ISABELLE_HOME_USER", "")
blanchet@49331
   315
  else (getenv "ISABELLE_TMP", serial_string ())
blanchet@49331
   316
blanchet@49347
   317
fun run_mash ctxt overlord (temp_dir, serial) core =
blanchet@49326
   318
  let
blanchet@49331
   319
    val log_file = temp_dir ^ "/mash_log" ^ serial
blanchet@49331
   320
    val err_file = temp_dir ^ "/mash_err" ^ serial
blanchet@49326
   321
    val command =
blanchet@49331
   322
      mash_home () ^ "/mash.py --quiet --outputDir " ^ mash_state_dir () ^
blanchet@49331
   323
      " --log " ^ log_file ^ " " ^ core ^ " 2>&1 > " ^ err_file
blanchet@49331
   324
  in
blanchet@49331
   325
    trace_msg ctxt (fn () => "Running " ^ command);
blanchet@49338
   326
    write_file ([], K "") log_file;
blanchet@49338
   327
    write_file ([], K "") err_file;
blanchet@49347
   328
    Isabelle_System.bash command;
blanchet@49347
   329
    if overlord then ()
blanchet@49347
   330
    else (map (File.rm o Path.explode) [log_file, err_file]; ());
blanchet@49347
   331
    trace_msg ctxt (K "Done")
blanchet@49331
   332
  end
blanchet@49326
   333
blanchet@49347
   334
(* TODO: Eliminate code once "mash.py" handles sequences of ADD commands as fast
blanchet@49347
   335
   as a single INIT. *)
blanchet@49331
   336
fun run_mash_init ctxt overlord write_access write_feats write_deps =
blanchet@49331
   337
  let
blanchet@49331
   338
    val info as (temp_dir, serial) = mash_info overlord
blanchet@49331
   339
    val in_dir = temp_dir ^ "/mash_init" ^ serial
blanchet@49347
   340
    val in_dir_path = in_dir |> Path.explode |> tap Isabelle_System.mkdir
blanchet@49331
   341
  in
blanchet@49331
   342
    write_file write_access (in_dir ^ "/mash_accessibility");
blanchet@49331
   343
    write_file write_feats (in_dir ^ "/mash_features");
blanchet@49331
   344
    write_file write_deps (in_dir ^ "/mash_dependencies");
blanchet@49347
   345
    run_mash ctxt overlord info ("--init --inputDir " ^ in_dir);
blanchet@49347
   346
    (* FIXME: temporary hack *)
blanchet@49347
   347
    if overlord then ()
blanchet@49347
   348
    else (Isabelle_System.bash ("rm -r -f " ^ File.shell_path in_dir_path); ())
blanchet@49331
   349
  end
blanchet@49331
   350
blanchet@49333
   351
fun run_mash_commands ctxt overlord save max_suggs write_cmds read_suggs =
blanchet@49331
   352
  let
blanchet@49331
   353
    val info as (temp_dir, serial) = mash_info overlord
blanchet@49333
   354
    val sugg_file = temp_dir ^ "/mash_suggs" ^ serial
blanchet@49331
   355
    val cmd_file = temp_dir ^ "/mash_commands" ^ serial
blanchet@49331
   356
  in
blanchet@49338
   357
    write_file ([], K "") sugg_file;
blanchet@49331
   358
    write_file write_cmds cmd_file;
blanchet@49347
   359
    run_mash ctxt overlord info
blanchet@49333
   360
             ("--inputFile " ^ cmd_file ^ " --predictions " ^ sugg_file ^
blanchet@49333
   361
              " --numberOfPredictions " ^ string_of_int max_suggs ^
blanchet@49331
   362
              (if save then " --saveModel" else ""));
blanchet@49333
   363
    read_suggs (fn () => File.read_lines (Path.explode sugg_file))
blanchet@49347
   364
    |> tap (fn _ =>
blanchet@49347
   365
               if overlord then ()
blanchet@49347
   366
               else (map (File.rm o Path.explode) [sugg_file, cmd_file]; ()))
blanchet@49331
   367
  end
blanchet@49331
   368
blanchet@49331
   369
fun str_of_update (name, parents, feats, deps) =
blanchet@49331
   370
  "! " ^ escape_meta name ^ ": " ^ escape_metas parents ^ "; " ^
blanchet@49326
   371
  escape_metas feats ^ "; " ^ escape_metas deps ^ "\n"
blanchet@49326
   372
blanchet@49331
   373
fun str_of_query (parents, feats) =
blanchet@49331
   374
  "? " ^ escape_metas parents ^ "; " ^ escape_metas feats
blanchet@49331
   375
blanchet@49331
   376
fun init_str_of_update get (upd as (name, _, _, _)) =
blanchet@49331
   377
  escape_meta name ^ ": " ^ escape_metas (get upd) ^ "\n"
blanchet@49326
   378
blanchet@49347
   379
fun mash_CLEAR ctxt =
blanchet@49329
   380
  let val path = mash_state_dir () |> Path.explode in
blanchet@49347
   381
    trace_msg ctxt (K "MaSh CLEAR");
blanchet@49324
   382
    File.fold_dir (fn file => fn () =>
blanchet@49324
   383
                      File.rm (Path.append path (Path.basic file)))
blanchet@49324
   384
                  path ()
blanchet@49324
   385
  end
blanchet@49317
   386
blanchet@49347
   387
fun mash_INIT ctxt _ [] = mash_CLEAR ctxt
blanchet@49331
   388
  | mash_INIT ctxt overlord upds =
blanchet@49331
   389
    (trace_msg ctxt (fn () => "MaSh INIT " ^
blanchet@49331
   390
         elide_string 1000 (space_implode " " (map #1 upds)));
blanchet@49338
   391
     run_mash_init ctxt overlord (upds, init_str_of_update #2)
blanchet@49338
   392
                   (upds, init_str_of_update #3) (upds, init_str_of_update #4))
blanchet@49317
   393
blanchet@49331
   394
fun mash_ADD _ _ [] = ()
blanchet@49331
   395
  | mash_ADD ctxt overlord upds =
blanchet@49331
   396
    (trace_msg ctxt (fn () => "MaSh ADD " ^
blanchet@49331
   397
         elide_string 1000 (space_implode " " (map #1 upds)));
blanchet@49338
   398
     run_mash_commands ctxt overlord true 0 (upds, str_of_update) (K ()))
blanchet@49317
   399
blanchet@49333
   400
fun mash_QUERY ctxt overlord max_suggs (query as (_, feats)) =
blanchet@49329
   401
  (trace_msg ctxt (fn () => "MaSh QUERY " ^ space_implode " " feats);
blanchet@49333
   402
   run_mash_commands ctxt overlord false max_suggs
blanchet@49338
   403
       ([query], str_of_query)
blanchet@49333
   404
       (fn suggs => snd (extract_query (List.last (suggs ()))))
blanchet@49326
   405
   handle List.Empty => [])
blanchet@49317
   406
blanchet@49317
   407
blanchet@49266
   408
(*** High-level communication with MaSh ***)
blanchet@49266
   409
blanchet@49336
   410
fun try_graph ctxt when def f =
blanchet@49336
   411
  f ()
blanchet@49336
   412
  handle Graph.CYCLES (cycle :: _) =>
blanchet@49336
   413
         (trace_msg ctxt (fn () =>
blanchet@49336
   414
              "Cycle involving " ^ commas cycle ^ " when " ^ when); def)
blanchet@49336
   415
       | Graph.UNDEF name =>
blanchet@49336
   416
         (trace_msg ctxt (fn () =>
blanchet@49336
   417
              "Unknown fact " ^ quote name ^ " when " ^ when); def)
blanchet@49336
   418
blanchet@49316
   419
type mash_state =
blanchet@49331
   420
  {thys : bool Symtab.table,
blanchet@49331
   421
   fact_graph : unit Graph.T}
blanchet@49264
   422
blanchet@49331
   423
val empty_state = {thys = Symtab.empty, fact_graph = Graph.empty}
blanchet@49316
   424
blanchet@49316
   425
local
blanchet@49316
   426
blanchet@49336
   427
fun mash_load _ (state as (true, _)) = state
blanchet@49336
   428
  | mash_load ctxt _ =
blanchet@49324
   429
    let val path = mash_state_path () in
blanchet@49317
   430
      (true,
blanchet@49317
   431
       case try File.read_lines path of
blanchet@49331
   432
         SOME (comp_thys :: incomp_thys :: fact_lines) =>
blanchet@49317
   433
         let
blanchet@49331
   434
           fun add_thy comp thy = Symtab.update (thy, comp)
blanchet@49337
   435
           fun add_edge_to name parent =
blanchet@49337
   436
             Graph.default_node (parent, ())
blanchet@49337
   437
             #> Graph.add_edge (parent, name)
blanchet@49331
   438
           fun add_fact_line line =
blanchet@49331
   439
             case extract_query line of
blanchet@49331
   440
               ("", _) => I (* shouldn't happen *)
blanchet@49331
   441
             | (name, parents) =>
blanchet@49331
   442
               Graph.default_node (name, ())
blanchet@49337
   443
               #> fold (add_edge_to name) parents
blanchet@49331
   444
           val thys =
blanchet@49331
   445
             Symtab.empty |> fold (add_thy true) (unescape_metas comp_thys)
blanchet@49331
   446
                          |> fold (add_thy false) (unescape_metas incomp_thys)
blanchet@49336
   447
           val fact_graph =
blanchet@49337
   448
             try_graph ctxt "loading state" Graph.empty (fn () =>
blanchet@49336
   449
                 Graph.empty |> fold add_fact_line fact_lines)
blanchet@49331
   450
         in {thys = thys, fact_graph = fact_graph} end
blanchet@49319
   451
       | _ => empty_state)
blanchet@49317
   452
    end
blanchet@49316
   453
blanchet@49331
   454
fun mash_save ({thys, fact_graph, ...} : mash_state) =
blanchet@49316
   455
  let
blanchet@49324
   456
    val path = mash_state_path ()
blanchet@49331
   457
    val thys = Symtab.dest thys
blanchet@49333
   458
    val line_for_thys = escape_metas o AList.find (op =) thys
blanchet@49333
   459
    fun fact_line_for name parents =
blanchet@49333
   460
      escape_meta name ^ ": " ^ escape_metas parents
blanchet@49331
   461
    val append_fact = File.append path o suffix "\n" oo fact_line_for
blanchet@49316
   462
  in
blanchet@49331
   463
    File.write path (line_for_thys true ^ "\n" ^ line_for_thys false ^ "\n");
blanchet@49331
   464
    Graph.fold (fn (name, ((), (parents, _))) => fn () =>
blanchet@49331
   465
                   append_fact name (Graph.Keys.dest parents))
blanchet@49331
   466
        fact_graph ()
blanchet@49316
   467
  end
blanchet@49316
   468
blanchet@49317
   469
val global_state =
blanchet@49319
   470
  Synchronized.var "Sledgehammer_Filter_MaSh.global_state" (false, empty_state)
blanchet@49316
   471
blanchet@49316
   472
in
blanchet@49316
   473
blanchet@49336
   474
fun mash_map ctxt f =
blanchet@49336
   475
  Synchronized.change global_state (mash_load ctxt ##> (f #> tap mash_save))
blanchet@49316
   476
blanchet@49336
   477
fun mash_get ctxt =
blanchet@49336
   478
  Synchronized.change_result global_state (mash_load ctxt #> `snd)
blanchet@49317
   479
blanchet@49347
   480
fun mash_unlearn ctxt =
blanchet@49317
   481
  Synchronized.change global_state (fn _ =>
blanchet@49347
   482
      (mash_CLEAR ctxt; File.write (mash_state_path ()) "";
blanchet@49323
   483
       (true, empty_state)))
blanchet@49316
   484
blanchet@49316
   485
end
blanchet@49316
   486
blanchet@49333
   487
fun mash_could_suggest_facts () = mash_home () <> ""
blanchet@49336
   488
fun mash_can_suggest_facts ctxt =
blanchet@49336
   489
  not (Graph.is_empty (#fact_graph (mash_get ctxt)))
blanchet@49264
   490
blanchet@49347
   491
fun parents_wrt_facts facts fact_graph =
blanchet@49331
   492
  let
blanchet@49393
   493
    val facts = [] |> fold (cons o nickname_of o snd) facts
blanchet@49345
   494
    val tab = Symtab.empty |> fold (fn name => Symtab.update (name, ())) facts
blanchet@49392
   495
    fun insert_not_seen seen name =
blanchet@49392
   496
      not (member (op =) seen name) ? insert (op =) name
blanchet@49393
   497
    fun parents_of _ parents [] = parents
blanchet@49392
   498
      | parents_of seen parents (name :: names) =
blanchet@49347
   499
        if Symtab.defined tab name then
blanchet@49392
   500
          parents_of (name :: seen) (name :: parents) names
blanchet@49347
   501
        else
blanchet@49392
   502
          parents_of (name :: seen) parents
blanchet@49392
   503
                     (Graph.Keys.fold (insert_not_seen seen)
blanchet@49392
   504
                                      (Graph.imm_preds fact_graph name) names)
blanchet@49392
   505
  in parents_of [] [] (Graph.maximals fact_graph) end
blanchet@49331
   506
blanchet@49333
   507
(* Generate more suggestions than requested, because some might be thrown out
blanchet@49333
   508
   later for various reasons and "meshing" gives better results with some
blanchet@49333
   509
   slack. *)
blanchet@49333
   510
fun max_suggs_of max_facts = max_facts + Int.min (200, max_facts)
blanchet@49333
   511
blanchet@49335
   512
fun is_fact_in_graph fact_graph (_, th) =
blanchet@49393
   513
  can (Graph.get_node fact_graph) (nickname_of th)
blanchet@49335
   514
blanchet@49333
   515
fun mash_suggest_facts ctxt ({overlord, ...} : params) prover max_facts hyp_ts
blanchet@49333
   516
                       concl_t facts =
blanchet@49316
   517
  let
blanchet@49317
   518
    val thy = Proof_Context.theory_of ctxt
blanchet@49336
   519
    val fact_graph = #fact_graph (mash_get ctxt)
blanchet@49347
   520
    val parents = parents_wrt_facts facts fact_graph
blanchet@49333
   521
    val feats = features_of ctxt prover thy General (concl_t :: hyp_ts)
blanchet@49333
   522
    val suggs =
blanchet@49335
   523
      if Graph.is_empty fact_graph then []
blanchet@49335
   524
      else mash_QUERY ctxt overlord (max_suggs_of max_facts) (parents, feats)
blanchet@49335
   525
    val selected = facts |> suggested_facts suggs
blanchet@49335
   526
    val unknown = facts |> filter_out (is_fact_in_graph fact_graph)
blanchet@49335
   527
  in (selected, unknown) end
blanchet@49264
   528
blanchet@49331
   529
fun add_thys_for thy =
blanchet@49331
   530
  let fun add comp thy = Symtab.update (Context.theory_name thy, comp) in
blanchet@49331
   531
    add false thy #> fold (add true) (Theory.ancestors_of thy)
blanchet@49331
   532
  end
blanchet@49319
   533
blanchet@49331
   534
fun update_fact_graph ctxt (name, parents, feats, deps) (upds, graph) =
blanchet@49331
   535
  let
blanchet@49331
   536
    fun maybe_add_from from (accum as (parents, graph)) =
blanchet@49336
   537
      try_graph ctxt "updating graph" accum (fn () =>
blanchet@49336
   538
          (from :: parents, Graph.add_edge_acyclic (from, name) graph))
blanchet@49336
   539
    val graph = graph |> Graph.default_node (name, ())
blanchet@49331
   540
    val (parents, graph) = ([], graph) |> fold maybe_add_from parents
blanchet@49331
   541
    val (deps, graph) = ([], graph) |> fold maybe_add_from deps
blanchet@49331
   542
  in ((name, parents, feats, deps) :: upds, graph) end
blanchet@49321
   543
blanchet@49333
   544
val pass1_learn_timeout_factor = 0.5
blanchet@49333
   545
blanchet@49347
   546
(* Too many dependencies is a sign that a decision procedure is at work. There
blanchet@49347
   547
   isn't much too learn from such proofs. *)
blanchet@49347
   548
val max_dependencies = 10
blanchet@49347
   549
blanchet@49333
   550
(* The timeout is understood in a very slack fashion. *)
blanchet@49334
   551
fun mash_learn_thy ctxt ({provers, verbose, overlord, ...} : params) thy timeout
blanchet@49334
   552
                   facts =
blanchet@49319
   553
  let
blanchet@49333
   554
    val timer = Timer.startRealTimer ()
blanchet@49333
   555
    val prover = hd provers
blanchet@49333
   556
    fun timed_out frac =
blanchet@49333
   557
      Time.> (Timer.checkRealTimer timer, time_mult frac timeout)
blanchet@49336
   558
    val {fact_graph, ...} = mash_get ctxt
blanchet@49335
   559
    val new_facts =
blanchet@49335
   560
      facts |> filter_out (is_fact_in_graph fact_graph)
blanchet@49335
   561
            |> sort (thm_ord o pairself snd)
blanchet@49323
   562
  in
blanchet@49323
   563
    if null new_facts then
blanchet@49334
   564
      ""
blanchet@49323
   565
    else
blanchet@49319
   566
      let
blanchet@49323
   567
        val ths = facts |> map snd
blanchet@49330
   568
        val all_names =
blanchet@49339
   569
          ths |> filter_out is_likely_tautology_or_too_meta
blanchet@49393
   570
              |> map (rpair () o nickname_of)
blanchet@49331
   571
              |> Symtab.make
blanchet@49347
   572
        fun trim_deps deps = if length deps > max_dependencies then [] else deps
blanchet@49333
   573
        fun do_fact _ (accum as (_, true)) = accum
blanchet@49333
   574
          | do_fact ((_, (_, status)), th) ((parents, upds), false) =
blanchet@49333
   575
            let
blanchet@49393
   576
              val name = nickname_of th
blanchet@49347
   577
              val feats =
blanchet@49347
   578
                features_of ctxt prover (theory_of_thm th) status [prop_of th]
blanchet@49347
   579
              val deps = isabelle_dependencies_of all_names th |> trim_deps
blanchet@49333
   580
              val upd = (name, parents, feats, deps)
blanchet@49333
   581
            in (([name], upd :: upds), timed_out pass1_learn_timeout_factor) end
blanchet@49347
   582
        val parents = parents_wrt_facts facts fact_graph
blanchet@49333
   583
        val ((_, upds), _) =
blanchet@49333
   584
          ((parents, []), false) |> fold do_fact new_facts |>> apsnd rev
blanchet@49333
   585
        val n = length upds
blanchet@49331
   586
        fun trans {thys, fact_graph} =
blanchet@49331
   587
          let
blanchet@49331
   588
            val mash_INIT_or_ADD =
blanchet@49331
   589
              if Graph.is_empty fact_graph then mash_INIT else mash_ADD
blanchet@49331
   590
            val (upds, fact_graph) =
blanchet@49331
   591
              ([], fact_graph) |> fold (update_fact_graph ctxt) upds
blanchet@49331
   592
          in
blanchet@49331
   593
            (mash_INIT_or_ADD ctxt overlord (rev upds);
blanchet@49331
   594
             {thys = thys |> add_thys_for thy,
blanchet@49331
   595
              fact_graph = fact_graph})
blanchet@49331
   596
          end
blanchet@49333
   597
      in
blanchet@49336
   598
        mash_map ctxt trans;
blanchet@49334
   599
        if verbose then
blanchet@49334
   600
          "Processed " ^ string_of_int n ^ " proof" ^ plural_s n ^
blanchet@49334
   601
          (if verbose then
blanchet@49334
   602
             " in " ^ string_from_time (Timer.checkRealTimer timer)
blanchet@49334
   603
           else
blanchet@49334
   604
             "") ^ "."
blanchet@49334
   605
        else
blanchet@49334
   606
          ""
blanchet@49333
   607
      end
blanchet@49323
   608
  end
blanchet@49319
   609
blanchet@49336
   610
fun mash_learn_proof ctxt ({provers, overlord, ...} : params) t facts used_ths =
blanchet@49331
   611
  let
blanchet@49331
   612
    val thy = Proof_Context.theory_of ctxt
blanchet@49333
   613
    val prover = hd provers
blanchet@49342
   614
    val name = ATP_Util.timestamp () ^ " " ^ serial_string () (* fresh enough *)
blanchet@49333
   615
    val feats = features_of ctxt prover thy General [t]
blanchet@49393
   616
    val deps = used_ths |> map nickname_of
blanchet@49331
   617
  in
blanchet@49336
   618
    mash_map ctxt (fn {thys, fact_graph} =>
blanchet@49324
   619
        let
blanchet@49347
   620
          val parents = parents_wrt_facts facts fact_graph
blanchet@49331
   621
          val upds = [(name, parents, feats, deps)]
blanchet@49331
   622
          val (upds, fact_graph) =
blanchet@49331
   623
            ([], fact_graph) |> fold (update_fact_graph ctxt) upds
blanchet@49324
   624
        in
blanchet@49331
   625
          mash_ADD ctxt overlord upds;
blanchet@49331
   626
          {thys = thys, fact_graph = fact_graph}
blanchet@49331
   627
        end)
blanchet@49331
   628
  end
blanchet@49264
   629
blanchet@49333
   630
(* The threshold should be large enough so that MaSh doesn't kick in for Auto
blanchet@49333
   631
   Sledgehammer and Try. *)
blanchet@49333
   632
val min_secs_for_learning = 15
blanchet@49342
   633
val learn_timeout_factor = 2.0
blanchet@49333
   634
blanchet@49336
   635
fun relevant_facts ctxt (params as {learn, fact_filter, timeout, ...}) prover
blanchet@49336
   636
        max_facts ({add, only, ...} : fact_override) hyp_ts concl_t facts =
blanchet@49329
   637
  if not (subset (op =) (the_list fact_filter, fact_filters)) then
blanchet@49329
   638
    error ("Unknown fact filter: " ^ quote (the fact_filter) ^ ".")
blanchet@49329
   639
  else if only then
blanchet@49304
   640
    facts
blanchet@49336
   641
  else if max_facts <= 0 orelse null facts then
blanchet@49303
   642
    []
blanchet@49303
   643
  else
blanchet@49303
   644
    let
blanchet@49333
   645
      val thy = Proof_Context.theory_of ctxt
blanchet@49342
   646
      fun maybe_learn () =
blanchet@49336
   647
        if not learn orelse Async_Manager.has_running_threads MaShN then
blanchet@49334
   648
          ()
blanchet@49334
   649
        else if Time.toSeconds timeout >= min_secs_for_learning then
blanchet@49334
   650
          let
blanchet@49342
   651
            val soft_timeout = time_mult learn_timeout_factor timeout
blanchet@49342
   652
            val hard_timeout = time_mult 4.0 soft_timeout
blanchet@49334
   653
            val birth_time = Time.now ()
blanchet@49334
   654
            val death_time = Time.+ (birth_time, hard_timeout)
blanchet@49334
   655
            val desc = ("machine learner for Sledgehammer", "")
blanchet@49334
   656
          in
blanchet@49334
   657
            Async_Manager.launch MaShN birth_time death_time desc
blanchet@49334
   658
                (fn () =>
blanchet@49334
   659
                    (true, mash_learn_thy ctxt params thy soft_timeout facts))
blanchet@49334
   660
          end
blanchet@49333
   661
        else
blanchet@49333
   662
          ()
blanchet@49329
   663
      val fact_filter =
blanchet@49329
   664
        case fact_filter of
blanchet@49342
   665
          SOME ff => (() |> ff <> iterN ? maybe_learn; ff)
blanchet@49333
   666
        | NONE =>
blanchet@49342
   667
          if mash_can_suggest_facts ctxt then (maybe_learn (); meshN)
blanchet@49342
   668
          else if mash_could_suggest_facts () then (maybe_learn (); iterN)
blanchet@49333
   669
          else iterN
blanchet@49303
   670
      val add_ths = Attrib.eval_thms ctxt add
blanchet@49307
   671
      fun prepend_facts ths accepts =
blanchet@49303
   672
        ((facts |> filter (member Thm.eq_thm_prop ths o snd)) @
blanchet@49307
   673
         (accepts |> filter_out (member Thm.eq_thm_prop ths o snd)))
blanchet@49308
   674
        |> take max_facts
blanchet@49329
   675
      fun iter () =
blanchet@49313
   676
        iterative_relevant_facts ctxt params prover max_facts NONE hyp_ts
blanchet@49313
   677
                                 concl_t facts
blanchet@49329
   678
      fun mash () =
blanchet@49335
   679
        mash_suggest_facts ctxt params prover max_facts hyp_ts concl_t facts
blanchet@49329
   680
      val mess =
blanchet@49335
   681
        [] |> (if fact_filter <> mashN then cons (iter (), []) else I)
blanchet@49329
   682
           |> (if fact_filter <> iterN then cons (mash ()) else I)
blanchet@49303
   683
    in
blanchet@49328
   684
      mesh_facts max_facts mess
blanchet@49303
   685
      |> not (null add_ths) ? prepend_facts add_ths
blanchet@49303
   686
    end
blanchet@49303
   687
blanchet@49334
   688
fun kill_learners () = Async_Manager.kill_threads MaShN "learner"
blanchet@49334
   689
fun running_learners () = Async_Manager.running_threads MaShN "learner"
blanchet@49334
   690
blanchet@49263
   691
end;