src/HOL/Tools/Sledgehammer/sledgehammer_filter_mash.ML
author blanchet
Wed, 18 Jul 2012 08:44:04 +0200
changeset 49335 891a24a48155
parent 49334 340187063d84
child 49336 c552d7f1720b
permissions -rw-r--r--
improved meshing of MaSh and Meng--Paulson if some MaSh suggestions are cut-off (the common case)
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@49327
    28
  val suggested_facts : string list -> fact list -> fact list
blanchet@49335
    29
  val mesh_facts : int -> (fact list * fact list) list -> fact list
blanchet@49333
    30
  val is_likely_tautology : Proof.context -> string -> thm -> bool
blanchet@49330
    31
  val is_too_meta : thm -> bool
blanchet@49266
    32
  val theory_ord : theory * theory -> order
blanchet@49266
    33
  val thm_ord : thm * thm -> order
blanchet@49333
    34
  val features_of :
blanchet@49333
    35
    Proof.context -> string -> theory -> status -> term list -> string list
blanchet@49331
    36
  val isabelle_dependencies_of : unit Symtab.table -> thm -> string list
blanchet@49266
    37
  val goal_of_thm : theory -> thm -> thm
blanchet@49333
    38
  val run_prover :
blanchet@49333
    39
    Proof.context -> params -> string -> fact list -> thm -> prover_result
blanchet@49323
    40
  val mash_RESET : Proof.context -> unit
blanchet@49331
    41
  val mash_INIT :
blanchet@49331
    42
    Proof.context -> bool
blanchet@49331
    43
    -> (string * string list * string list * string list) list -> unit
blanchet@49323
    44
  val mash_ADD :
blanchet@49331
    45
    Proof.context -> bool
blanchet@49331
    46
    -> (string * string list * string list * string list) list -> unit
blanchet@49331
    47
  val mash_QUERY :
blanchet@49333
    48
    Proof.context -> bool -> int -> string list * string list -> string list
blanchet@49323
    49
  val mash_reset : Proof.context -> unit
blanchet@49333
    50
  val mash_could_suggest_facts : unit -> bool
blanchet@49333
    51
  val mash_can_suggest_facts : unit -> bool
blanchet@49313
    52
  val mash_suggest_facts :
blanchet@49333
    53
    Proof.context -> params -> string -> int -> term list -> term -> fact list
blanchet@49335
    54
    -> fact list * fact list
blanchet@49334
    55
  val mash_learn_thy :
blanchet@49334
    56
    Proof.context -> params -> theory -> Time.time -> fact list -> string
blanchet@49331
    57
  val mash_learn_proof :
blanchet@49331
    58
    Proof.context -> params -> term -> thm list -> fact list -> unit
blanchet@49303
    59
  val relevant_facts :
blanchet@49307
    60
    Proof.context -> params -> string -> int -> fact_override -> term list
blanchet@49311
    61
    -> term -> fact list -> fact list
blanchet@49334
    62
  val kill_learners : unit -> unit
blanchet@49334
    63
  val running_learners : unit -> unit
blanchet@49263
    64
end;
blanchet@49263
    65
blanchet@49263
    66
structure Sledgehammer_Filter_MaSh : SLEDGEHAMMER_FILTER_MASH =
blanchet@49263
    67
struct
blanchet@49264
    68
blanchet@49266
    69
open ATP_Util
blanchet@49266
    70
open ATP_Problem_Generate
blanchet@49266
    71
open Sledgehammer_Util
blanchet@49266
    72
open Sledgehammer_Fact
blanchet@49266
    73
open Sledgehammer_Filter_Iter
blanchet@49266
    74
open Sledgehammer_Provers
blanchet@49333
    75
open Sledgehammer_Minimize
blanchet@49266
    76
blanchet@49323
    77
val trace =
blanchet@49323
    78
  Attrib.setup_config_bool @{binding sledgehammer_filter_mash_trace} (K false)
blanchet@49323
    79
fun trace_msg ctxt msg = if Config.get ctxt trace then tracing (msg ()) else ()
blanchet@49323
    80
blanchet@49334
    81
val MaShN = "MaSh"
blanchet@49334
    82
blanchet@49329
    83
val meshN = "mesh"
blanchet@49329
    84
val iterN = "iter"
blanchet@49329
    85
val mashN = "mash"
blanchet@49329
    86
blanchet@49329
    87
val fact_filters = [meshN, iterN, mashN]
blanchet@49329
    88
blanchet@49329
    89
fun mash_home () = getenv "MASH_HOME"
blanchet@49329
    90
fun mash_state_dir () =
blanchet@49324
    91
  getenv "ISABELLE_HOME_USER" ^ "/mash"
blanchet@49331
    92
  |> tap (Isabelle_System.mkdir o Path.explode)
blanchet@49329
    93
fun mash_state_path () = mash_state_dir () ^ "/state" |> Path.explode
blanchet@49266
    94
blanchet@49266
    95
(*** Isabelle helpers ***)
blanchet@49266
    96
blanchet@49323
    97
fun meta_char c =
blanchet@49266
    98
  if Char.isAlphaNum c orelse c = #"_" orelse c = #"." orelse c = #"(" orelse
blanchet@49266
    99
     c = #")" orelse c = #"," then
blanchet@49266
   100
    String.str c
blanchet@49266
   101
  else
blanchet@49266
   102
    (* fixed width, in case more digits follow *)
blanchet@49266
   103
    "\\" ^ stringN_of_int 3 (Char.ord c)
blanchet@49266
   104
blanchet@49323
   105
fun unmeta_chars accum [] = String.implode (rev accum)
blanchet@49323
   106
  | unmeta_chars accum (#"\\" :: d1 :: d2 :: d3 :: cs) =
blanchet@49323
   107
    (case Int.fromString (String.implode [d1, d2, d3]) of
blanchet@49323
   108
       SOME n => unmeta_chars (Char.chr n :: accum) cs
blanchet@49323
   109
     | NONE => "" (* error *))
blanchet@49323
   110
  | unmeta_chars _ (#"\\" :: _) = "" (* error *)
blanchet@49323
   111
  | unmeta_chars accum (c :: cs) = unmeta_chars (c :: accum) cs
blanchet@49323
   112
blanchet@49323
   113
val escape_meta = String.translate meta_char
blanchet@49318
   114
val escape_metas = map escape_meta #> space_implode " "
blanchet@49330
   115
val unescape_meta = String.explode #> unmeta_chars []
blanchet@49330
   116
val unescape_metas =
blanchet@49330
   117
  space_explode " " #> filter_out (curry (op =) "") #> map unescape_meta
blanchet@49266
   118
blanchet@49326
   119
fun extract_query line =
blanchet@49326
   120
  case space_explode ":" line of
blanchet@49330
   121
    [goal_name, suggs] => (unescape_meta goal_name, unescape_metas suggs)
blanchet@49327
   122
  | _ => ("", [])
blanchet@49326
   123
blanchet@49326
   124
fun find_suggested facts sugg =
blanchet@49326
   125
  find_first (fn (_, th) => Thm.get_name_hint th = sugg) facts
blanchet@49326
   126
fun suggested_facts suggs facts = map_filter (find_suggested facts) suggs
blanchet@49326
   127
blanchet@49335
   128
fun sum_avg _ [] = 1000000000 (* big number *)
blanchet@49335
   129
  | sum_avg n xs = fold (Integer.add o Integer.mult n) xs 0 div (length xs)
blanchet@49328
   130
blanchet@49335
   131
fun mesh_facts max_facts [(selected, unknown)] =
blanchet@49335
   132
    take max_facts selected @ take (max_facts - length selected) unknown
blanchet@49329
   133
  | mesh_facts max_facts mess =
blanchet@49329
   134
    let
blanchet@49335
   135
      val mess = mess |> map (apfst (`length))
blanchet@49329
   136
      val n = length mess
blanchet@49329
   137
      val fact_eq = Thm.eq_thm o pairself snd
blanchet@49335
   138
      fun score_in fact ((sel_len, sels), unks) =
blanchet@49335
   139
        case find_index (curry fact_eq fact) sels of
blanchet@49335
   140
          ~1 => (case find_index (curry fact_eq fact) unks of
blanchet@49335
   141
                   ~1 => SOME sel_len
blanchet@49335
   142
                 | _ => NONE)
blanchet@49329
   143
        | j => SOME j
blanchet@49329
   144
      fun score_of fact = mess |> map_filter (score_in fact) |> sum_avg n
blanchet@49335
   145
      val facts = fold (union fact_eq o take max_facts o snd o fst) mess []
blanchet@49329
   146
    in
blanchet@49329
   147
      facts |> map (`score_of) |> sort (int_ord o pairself fst) |> map snd
blanchet@49329
   148
            |> take max_facts
blanchet@49329
   149
    end
blanchet@49327
   150
blanchet@49318
   151
val thy_feature_prefix = "y_"
blanchet@49266
   152
blanchet@49318
   153
val thy_feature_name_of = prefix thy_feature_prefix
blanchet@49318
   154
val const_name_of = prefix const_prefix
blanchet@49318
   155
val type_name_of = prefix type_const_prefix
blanchet@49318
   156
val class_name_of = prefix class_prefix
blanchet@49266
   157
blanchet@49266
   158
local
blanchet@49266
   159
blanchet@49266
   160
fun has_bool @{typ bool} = true
blanchet@49266
   161
  | has_bool (Type (_, Ts)) = exists has_bool Ts
blanchet@49266
   162
  | has_bool _ = false
blanchet@49266
   163
blanchet@49266
   164
fun has_fun (Type (@{type_name fun}, _)) = true
blanchet@49266
   165
  | has_fun (Type (_, Ts)) = exists has_fun Ts
blanchet@49266
   166
  | has_fun _ = false
blanchet@49266
   167
blanchet@49266
   168
val is_conn = member (op =)
blanchet@49266
   169
  [@{const_name Trueprop}, @{const_name HOL.conj}, @{const_name HOL.disj},
blanchet@49266
   170
   @{const_name HOL.implies}, @{const_name Not},
blanchet@49266
   171
   @{const_name All}, @{const_name Ex}, @{const_name Ball}, @{const_name Bex},
blanchet@49266
   172
   @{const_name HOL.eq}]
blanchet@49266
   173
blanchet@49266
   174
val has_bool_arg_const =
blanchet@49266
   175
  exists_Const (fn (c, T) =>
blanchet@49266
   176
                   not (is_conn c) andalso exists has_bool (binder_types T))
blanchet@49266
   177
blanchet@49331
   178
fun higher_inst_const thy (s, T) =
blanchet@49266
   179
  case binder_types T of
blanchet@49266
   180
    [] => false
blanchet@49331
   181
  | Ts => length (binder_types (Sign.the_const_type thy s)) <> length Ts
blanchet@49331
   182
  handle TYPE _ => false
blanchet@49266
   183
blanchet@49266
   184
val binders = [@{const_name All}, @{const_name Ex}]
blanchet@49266
   185
blanchet@49266
   186
in
blanchet@49266
   187
blanchet@49266
   188
fun is_fo_term thy t =
blanchet@49266
   189
  let
blanchet@49266
   190
    val t =
blanchet@49266
   191
      t |> Envir.beta_eta_contract
blanchet@49266
   192
        |> transform_elim_prop
blanchet@49266
   193
        |> Object_Logic.atomize_term thy
blanchet@49266
   194
  in
blanchet@49266
   195
    Term.is_first_order binders t andalso
blanchet@49266
   196
    not (exists_subterm (fn Var (_, T) => has_bool T orelse has_fun T
blanchet@49266
   197
                          | _ => false) t orelse
blanchet@49266
   198
         has_bool_arg_const t orelse exists_Const (higher_inst_const thy) t)
blanchet@49266
   199
  end
blanchet@49266
   200
blanchet@49266
   201
end
blanchet@49266
   202
blanchet@49333
   203
fun interesting_terms_types_and_classes ctxt prover term_max_depth
blanchet@49333
   204
                                        type_max_depth ts =
blanchet@49266
   205
  let
blanchet@49266
   206
    val bad_types = [@{type_name prop}, @{type_name bool}, @{type_name fun}]
blanchet@49333
   207
    fun is_bad_const (x as (s, _)) args =
blanchet@49333
   208
      member (op =) atp_logical_consts s orelse
blanchet@49333
   209
      fst (is_built_in_const_for_prover ctxt prover x args)
blanchet@49319
   210
    fun add_classes @{sort type} = I
blanchet@49319
   211
      | add_classes S = union (op =) (map class_name_of S)
blanchet@49266
   212
    fun do_add_type (Type (s, Ts)) =
blanchet@49266
   213
        (not (member (op =) bad_types s) ? insert (op =) (type_name_of s))
blanchet@49266
   214
        #> fold do_add_type Ts
blanchet@49319
   215
      | do_add_type (TFree (_, S)) = add_classes S
blanchet@49319
   216
      | do_add_type (TVar (_, S)) = add_classes S
blanchet@49266
   217
    fun add_type T = type_max_depth >= 0 ? do_add_type T
blanchet@49266
   218
    fun mk_app s args =
blanchet@49266
   219
      if member (op <>) args "" then s ^ "(" ^ space_implode "," args ^ ")"
blanchet@49266
   220
      else s
blanchet@49266
   221
    fun patternify ~1 _ = ""
blanchet@49266
   222
      | patternify depth t =
blanchet@49266
   223
        case strip_comb t of
blanchet@49266
   224
          (Const (s, _), args) =>
blanchet@49266
   225
          mk_app (const_name_of s) (map (patternify (depth - 1)) args)
blanchet@49266
   226
        | _ => ""
blanchet@49266
   227
    fun add_term_patterns ~1 _ = I
blanchet@49266
   228
      | add_term_patterns depth t =
blanchet@49266
   229
        insert (op =) (patternify depth t)
blanchet@49266
   230
        #> add_term_patterns (depth - 1) t
blanchet@49266
   231
    val add_term = add_term_patterns term_max_depth
blanchet@49266
   232
    fun add_patterns t =
blanchet@49266
   233
      let val (head, args) = strip_comb t in
blanchet@49266
   234
        (case head of
blanchet@49333
   235
           Const (x as (_, T)) =>
blanchet@49333
   236
           not (is_bad_const x args) ? (add_term t #> add_type T)
blanchet@49266
   237
         | Free (_, T) => add_type T
blanchet@49266
   238
         | Var (_, T) => add_type T
blanchet@49266
   239
         | Abs (_, T, body) => add_type T #> add_patterns body
blanchet@49266
   240
         | _ => I)
blanchet@49266
   241
        #> fold add_patterns args
blanchet@49266
   242
      end
blanchet@49317
   243
  in [] |> fold add_patterns ts |> sort string_ord end
blanchet@49266
   244
blanchet@49333
   245
fun is_likely_tautology ctxt prover th =
blanchet@49333
   246
  null (interesting_terms_types_and_classes ctxt prover 0 ~1 [prop_of th])
blanchet@49333
   247
  andalso not (Thm.eq_thm_prop (@{thm ext}, th))
blanchet@49266
   248
blanchet@49330
   249
(* ### FIXME: optimize *)
blanchet@49330
   250
fun is_too_meta th =
blanchet@49330
   251
  fastype_of (Object_Logic.atomize_term (theory_of_thm th) (prop_of th))
blanchet@49330
   252
  <> @{typ bool}
blanchet@49266
   253
blanchet@49266
   254
fun theory_ord p =
blanchet@49330
   255
  if Theory.eq_thy p then
blanchet@49330
   256
    EQUAL
blanchet@49330
   257
  else if Theory.subthy p then
blanchet@49330
   258
    LESS
blanchet@49330
   259
  else if Theory.subthy (swap p) then
blanchet@49330
   260
    GREATER
blanchet@49330
   261
  else case int_ord (pairself (length o Theory.ancestors_of) p) of
blanchet@49330
   262
    EQUAL => string_ord (pairself Context.theory_name p)
blanchet@49330
   263
  | order => order
blanchet@49266
   264
blanchet@49266
   265
val thm_ord = theory_ord o pairself theory_of_thm
blanchet@49266
   266
blanchet@49266
   267
fun is_exists (s, _) = (s = @{const_name Ex} orelse s = @{const_name Ex1})
blanchet@49266
   268
blanchet@49312
   269
val term_max_depth = 1
blanchet@49312
   270
val type_max_depth = 1
blanchet@49266
   271
blanchet@49266
   272
(* TODO: Generate type classes for types? *)
blanchet@49333
   273
fun features_of ctxt prover thy status ts =
blanchet@49318
   274
  thy_feature_name_of (Context.theory_name thy) ::
blanchet@49333
   275
  interesting_terms_types_and_classes ctxt prover term_max_depth type_max_depth
blanchet@49333
   276
                                      ts
blanchet@49317
   277
  |> exists (not o is_lambda_free) ts ? cons "lambdas"
blanchet@49317
   278
  |> exists (exists_Const is_exists) ts ? cons "skolems"
blanchet@49317
   279
  |> exists (not o is_fo_term thy) ts ? cons "ho"
blanchet@49317
   280
  |> (case status of
blanchet@49317
   281
        General => I
blanchet@49317
   282
      | Induction => cons "induction"
blanchet@49317
   283
      | Intro => cons "intro"
blanchet@49317
   284
      | Inductive => cons "inductive"
blanchet@49317
   285
      | Elim => cons "elim"
blanchet@49317
   286
      | Simp => cons "simp"
blanchet@49317
   287
      | Def => cons "def")
blanchet@49266
   288
blanchet@49266
   289
fun isabelle_dependencies_of all_facts =
blanchet@49318
   290
  thms_in_proof (SOME all_facts) #> sort string_ord
blanchet@49266
   291
blanchet@49266
   292
val freezeT = Type.legacy_freeze_type
blanchet@49266
   293
blanchet@49266
   294
fun freeze (t $ u) = freeze t $ freeze u
blanchet@49266
   295
  | freeze (Abs (s, T, t)) = Abs (s, freezeT T, freeze t)
blanchet@49266
   296
  | freeze (Var ((s, _), T)) = Free (s, freezeT T)
blanchet@49266
   297
  | freeze (Const (s, T)) = Const (s, freezeT T)
blanchet@49266
   298
  | freeze (Free (s, T)) = Free (s, freezeT T)
blanchet@49266
   299
  | freeze t = t
blanchet@49266
   300
blanchet@49266
   301
fun goal_of_thm thy = prop_of #> freeze #> cterm_of thy #> Goal.init
blanchet@49266
   302
blanchet@49333
   303
fun run_prover ctxt params prover facts goal =
blanchet@49266
   304
  let
blanchet@49266
   305
    val problem =
blanchet@49266
   306
      {state = Proof.init ctxt, goal = goal, subgoal = 1, subgoal_count = 1,
blanchet@49304
   307
       facts = facts |> map (apfst (apfst (fn name => name ())))
blanchet@49333
   308
                     |> map Untranslated_Fact}
blanchet@49333
   309
    val prover = get_minimizing_prover ctxt Normal prover
blanchet@49266
   310
  in prover params (K (K (K ""))) problem end
blanchet@49266
   311
blanchet@49266
   312
blanchet@49317
   313
(*** Low-level communication with MaSh ***)
blanchet@49317
   314
blanchet@49333
   315
fun write_file write file =
blanchet@49333
   316
  let val path = Path.explode file in
blanchet@49333
   317
    File.write path ""; write (File.append path)
blanchet@49333
   318
  end
blanchet@49331
   319
blanchet@49331
   320
fun mash_info overlord =
blanchet@49331
   321
  if overlord then (getenv "ISABELLE_HOME_USER", "")
blanchet@49331
   322
  else (getenv "ISABELLE_TMP", serial_string ())
blanchet@49331
   323
blanchet@49331
   324
fun run_mash ctxt (temp_dir, serial) core =
blanchet@49326
   325
  let
blanchet@49331
   326
    val log_file = temp_dir ^ "/mash_log" ^ serial
blanchet@49331
   327
    val err_file = temp_dir ^ "/mash_err" ^ serial
blanchet@49326
   328
    val command =
blanchet@49331
   329
      mash_home () ^ "/mash.py --quiet --outputDir " ^ mash_state_dir () ^
blanchet@49331
   330
      " --log " ^ log_file ^ " " ^ core ^ " 2>&1 > " ^ err_file
blanchet@49331
   331
  in
blanchet@49331
   332
    trace_msg ctxt (fn () => "Running " ^ command);
blanchet@49333
   333
    write_file (K ()) log_file;
blanchet@49333
   334
    write_file (K ()) err_file;
blanchet@49331
   335
    Isabelle_System.bash command; ()
blanchet@49331
   336
  end
blanchet@49326
   337
blanchet@49331
   338
fun run_mash_init ctxt overlord write_access write_feats write_deps =
blanchet@49331
   339
  let
blanchet@49331
   340
    val info as (temp_dir, serial) = mash_info overlord
blanchet@49331
   341
    val in_dir = temp_dir ^ "/mash_init" ^ serial
blanchet@49331
   342
                 |> tap (Isabelle_System.mkdir o Path.explode)
blanchet@49331
   343
  in
blanchet@49331
   344
    write_file write_access (in_dir ^ "/mash_accessibility");
blanchet@49331
   345
    write_file write_feats (in_dir ^ "/mash_features");
blanchet@49331
   346
    write_file write_deps (in_dir ^ "/mash_dependencies");
blanchet@49331
   347
    run_mash ctxt info ("--init --inputDir " ^ in_dir)
blanchet@49331
   348
  end
blanchet@49331
   349
blanchet@49333
   350
fun run_mash_commands ctxt overlord save max_suggs write_cmds read_suggs =
blanchet@49331
   351
  let
blanchet@49331
   352
    val info as (temp_dir, serial) = mash_info overlord
blanchet@49333
   353
    val sugg_file = temp_dir ^ "/mash_suggs" ^ serial
blanchet@49331
   354
    val cmd_file = temp_dir ^ "/mash_commands" ^ serial
blanchet@49331
   355
  in
blanchet@49333
   356
    write_file (K ()) sugg_file;
blanchet@49331
   357
    write_file write_cmds cmd_file;
blanchet@49331
   358
    run_mash ctxt info
blanchet@49333
   359
             ("--inputFile " ^ cmd_file ^ " --predictions " ^ sugg_file ^
blanchet@49333
   360
              " --numberOfPredictions " ^ string_of_int max_suggs ^
blanchet@49331
   361
              (if save then " --saveModel" else ""));
blanchet@49333
   362
    read_suggs (fn () => File.read_lines (Path.explode sugg_file))
blanchet@49331
   363
  end
blanchet@49331
   364
blanchet@49331
   365
fun str_of_update (name, parents, feats, deps) =
blanchet@49331
   366
  "! " ^ escape_meta name ^ ": " ^ escape_metas parents ^ "; " ^
blanchet@49326
   367
  escape_metas feats ^ "; " ^ escape_metas deps ^ "\n"
blanchet@49326
   368
blanchet@49331
   369
fun str_of_query (parents, feats) =
blanchet@49331
   370
  "? " ^ escape_metas parents ^ "; " ^ escape_metas feats
blanchet@49331
   371
blanchet@49331
   372
fun init_str_of_update get (upd as (name, _, _, _)) =
blanchet@49331
   373
  escape_meta name ^ ": " ^ escape_metas (get upd) ^ "\n"
blanchet@49326
   374
blanchet@49323
   375
fun mash_RESET ctxt =
blanchet@49329
   376
  let val path = mash_state_dir () |> Path.explode in
blanchet@49324
   377
    trace_msg ctxt (K "MaSh RESET");
blanchet@49324
   378
    File.fold_dir (fn file => fn () =>
blanchet@49324
   379
                      File.rm (Path.append path (Path.basic file)))
blanchet@49324
   380
                  path ()
blanchet@49324
   381
  end
blanchet@49317
   382
blanchet@49331
   383
fun mash_INIT ctxt _ [] = mash_RESET ctxt
blanchet@49331
   384
  | mash_INIT ctxt overlord upds =
blanchet@49331
   385
    (trace_msg ctxt (fn () => "MaSh INIT " ^
blanchet@49331
   386
         elide_string 1000 (space_implode " " (map #1 upds)));
blanchet@49331
   387
     run_mash_init ctxt overlord
blanchet@49331
   388
         (fn append => List.app (append o init_str_of_update #2) upds)
blanchet@49331
   389
         (fn append => List.app (append o init_str_of_update #3) upds)
blanchet@49331
   390
         (fn append => List.app (append o init_str_of_update #4) upds))
blanchet@49317
   391
blanchet@49331
   392
fun mash_ADD _ _ [] = ()
blanchet@49331
   393
  | mash_ADD ctxt overlord upds =
blanchet@49331
   394
    (trace_msg ctxt (fn () => "MaSh ADD " ^
blanchet@49331
   395
         elide_string 1000 (space_implode " " (map #1 upds)));
blanchet@49333
   396
     run_mash_commands ctxt overlord true 0
blanchet@49331
   397
         (fn append => List.app (append o str_of_update) upds) (K ()))
blanchet@49317
   398
blanchet@49333
   399
fun mash_QUERY ctxt overlord max_suggs (query as (_, feats)) =
blanchet@49329
   400
  (trace_msg ctxt (fn () => "MaSh QUERY " ^ space_implode " " feats);
blanchet@49333
   401
   run_mash_commands ctxt overlord false max_suggs
blanchet@49331
   402
       (fn append => append (str_of_query query))
blanchet@49333
   403
       (fn suggs => snd (extract_query (List.last (suggs ()))))
blanchet@49326
   404
   handle List.Empty => [])
blanchet@49317
   405
blanchet@49317
   406
blanchet@49266
   407
(*** High-level communication with MaSh ***)
blanchet@49266
   408
blanchet@49316
   409
type mash_state =
blanchet@49331
   410
  {thys : bool Symtab.table,
blanchet@49331
   411
   fact_graph : unit Graph.T}
blanchet@49264
   412
blanchet@49331
   413
val empty_state = {thys = Symtab.empty, fact_graph = Graph.empty}
blanchet@49316
   414
blanchet@49316
   415
local
blanchet@49316
   416
blanchet@49316
   417
fun mash_load (state as (true, _)) = state
blanchet@49316
   418
  | mash_load _ =
blanchet@49324
   419
    let val path = mash_state_path () in
blanchet@49317
   420
      (true,
blanchet@49317
   421
       case try File.read_lines path of
blanchet@49331
   422
         SOME (comp_thys :: incomp_thys :: fact_lines) =>
blanchet@49317
   423
         let
blanchet@49331
   424
           fun add_thy comp thy = Symtab.update (thy, comp)
blanchet@49331
   425
           fun add_fact_line line =
blanchet@49331
   426
             case extract_query line of
blanchet@49331
   427
               ("", _) => I (* shouldn't happen *)
blanchet@49331
   428
             | (name, parents) =>
blanchet@49331
   429
               Graph.default_node (name, ())
blanchet@49331
   430
               #> fold (fn par => Graph.add_edge (par, name)) parents
blanchet@49331
   431
           val thys =
blanchet@49331
   432
             Symtab.empty |> fold (add_thy true) (unescape_metas comp_thys)
blanchet@49331
   433
                          |> fold (add_thy false) (unescape_metas incomp_thys)
blanchet@49331
   434
           val fact_graph = Graph.empty |> fold add_fact_line fact_lines
blanchet@49331
   435
         in {thys = thys, fact_graph = fact_graph} end
blanchet@49319
   436
       | _ => empty_state)
blanchet@49317
   437
    end
blanchet@49316
   438
blanchet@49331
   439
fun mash_save ({thys, fact_graph, ...} : mash_state) =
blanchet@49316
   440
  let
blanchet@49324
   441
    val path = mash_state_path ()
blanchet@49331
   442
    val thys = Symtab.dest thys
blanchet@49333
   443
    val line_for_thys = escape_metas o AList.find (op =) thys
blanchet@49333
   444
    fun fact_line_for name parents =
blanchet@49333
   445
      escape_meta name ^ ": " ^ escape_metas parents
blanchet@49331
   446
    val append_fact = File.append path o suffix "\n" oo fact_line_for
blanchet@49316
   447
  in
blanchet@49331
   448
    File.write path (line_for_thys true ^ "\n" ^ line_for_thys false ^ "\n");
blanchet@49331
   449
    Graph.fold (fn (name, ((), (parents, _))) => fn () =>
blanchet@49331
   450
                   append_fact name (Graph.Keys.dest parents))
blanchet@49331
   451
        fact_graph ()
blanchet@49316
   452
  end
blanchet@49316
   453
blanchet@49317
   454
val global_state =
blanchet@49319
   455
  Synchronized.var "Sledgehammer_Filter_MaSh.global_state" (false, empty_state)
blanchet@49316
   456
blanchet@49316
   457
in
blanchet@49316
   458
blanchet@49321
   459
fun mash_map f =
blanchet@49317
   460
  Synchronized.change global_state (mash_load ##> (f #> tap mash_save))
blanchet@49316
   461
blanchet@49319
   462
fun mash_get () = Synchronized.change_result global_state (mash_load #> `snd)
blanchet@49317
   463
blanchet@49323
   464
fun mash_reset ctxt =
blanchet@49317
   465
  Synchronized.change global_state (fn _ =>
blanchet@49324
   466
      (mash_RESET ctxt; File.write (mash_state_path ()) "";
blanchet@49323
   467
       (true, empty_state)))
blanchet@49316
   468
blanchet@49316
   469
end
blanchet@49316
   470
blanchet@49333
   471
fun mash_could_suggest_facts () = mash_home () <> ""
blanchet@49333
   472
fun mash_can_suggest_facts () = not (Graph.is_empty (#fact_graph (mash_get ())))
blanchet@49264
   473
blanchet@49333
   474
fun parents_wrt_facts facts fact_graph =
blanchet@49331
   475
  let
blanchet@49333
   476
    val graph_facts = Symtab.make (map (rpair ()) (Graph.keys fact_graph))
blanchet@49331
   477
    val facts =
blanchet@49333
   478
      [] |> fold (cons o Thm.get_name_hint o snd) facts
blanchet@49333
   479
         |> filter (Symtab.defined graph_facts)
blanchet@49333
   480
         |> Graph.all_preds fact_graph
blanchet@49333
   481
    val facts = Symtab.empty |> fold (fn name => Symtab.update (name, ())) facts
blanchet@49331
   482
  in fact_graph |> Graph.restrict (Symtab.defined facts) |> Graph.maximals end
blanchet@49331
   483
blanchet@49333
   484
(* Generate more suggestions than requested, because some might be thrown out
blanchet@49333
   485
   later for various reasons and "meshing" gives better results with some
blanchet@49333
   486
   slack. *)
blanchet@49333
   487
fun max_suggs_of max_facts = max_facts + Int.min (200, max_facts)
blanchet@49333
   488
blanchet@49335
   489
fun is_fact_in_graph fact_graph (_, th) =
blanchet@49335
   490
  can (Graph.get_node fact_graph) (Thm.get_name_hint th)
blanchet@49335
   491
blanchet@49333
   492
fun mash_suggest_facts ctxt ({overlord, ...} : params) prover max_facts hyp_ts
blanchet@49333
   493
                       concl_t facts =
blanchet@49316
   494
  let
blanchet@49317
   495
    val thy = Proof_Context.theory_of ctxt
blanchet@49331
   496
    val fact_graph = #fact_graph (mash_get ())
blanchet@49333
   497
    val parents = parents_wrt_facts facts fact_graph
blanchet@49333
   498
    val feats = features_of ctxt prover thy General (concl_t :: hyp_ts)
blanchet@49333
   499
    val suggs =
blanchet@49335
   500
      if Graph.is_empty fact_graph then []
blanchet@49335
   501
      else mash_QUERY ctxt overlord (max_suggs_of max_facts) (parents, feats)
blanchet@49335
   502
    val selected = facts |> suggested_facts suggs
blanchet@49335
   503
    val unknown = facts |> filter_out (is_fact_in_graph fact_graph)
blanchet@49335
   504
  in (selected, unknown) end
blanchet@49264
   505
blanchet@49331
   506
fun add_thys_for thy =
blanchet@49331
   507
  let fun add comp thy = Symtab.update (Context.theory_name thy, comp) in
blanchet@49331
   508
    add false thy #> fold (add true) (Theory.ancestors_of thy)
blanchet@49331
   509
  end
blanchet@49319
   510
blanchet@49331
   511
fun update_fact_graph ctxt (name, parents, feats, deps) (upds, graph) =
blanchet@49331
   512
  let
blanchet@49331
   513
    fun maybe_add_from from (accum as (parents, graph)) =
blanchet@49331
   514
      (from :: parents, Graph.add_edge_acyclic (from, name) graph)
blanchet@49331
   515
      handle Graph.CYCLES _ =>
blanchet@49331
   516
             (trace_msg ctxt (fn () =>
blanchet@49331
   517
                  "Cycle between " ^ quote from ^ " and " ^ quote name); accum)
blanchet@49331
   518
           | Graph.UNDEF _ =>
blanchet@49331
   519
             (trace_msg ctxt (fn () => "Unknown node " ^ quote from); accum)
blanchet@49331
   520
    val graph = graph |> Graph.new_node (name, ())
blanchet@49331
   521
    val (parents, graph) = ([], graph) |> fold maybe_add_from parents
blanchet@49331
   522
    val (deps, graph) = ([], graph) |> fold maybe_add_from deps
blanchet@49331
   523
  in ((name, parents, feats, deps) :: upds, graph) end
blanchet@49321
   524
blanchet@49333
   525
val pass1_learn_timeout_factor = 0.5
blanchet@49333
   526
blanchet@49333
   527
(* The timeout is understood in a very slack fashion. *)
blanchet@49334
   528
fun mash_learn_thy ctxt ({provers, verbose, overlord, ...} : params) thy timeout
blanchet@49334
   529
                   facts =
blanchet@49319
   530
  let
blanchet@49333
   531
    val timer = Timer.startRealTimer ()
blanchet@49333
   532
    val prover = hd provers
blanchet@49333
   533
    fun timed_out frac =
blanchet@49333
   534
      Time.> (Timer.checkRealTimer timer, time_mult frac timeout)
blanchet@49331
   535
    val {fact_graph, ...} = mash_get ()
blanchet@49335
   536
    val new_facts =
blanchet@49335
   537
      facts |> filter_out (is_fact_in_graph fact_graph)
blanchet@49335
   538
            |> sort (thm_ord o pairself snd)
blanchet@49323
   539
  in
blanchet@49323
   540
    if null new_facts then
blanchet@49334
   541
      ""
blanchet@49323
   542
    else
blanchet@49319
   543
      let
blanchet@49323
   544
        val ths = facts |> map snd
blanchet@49330
   545
        val all_names =
blanchet@49333
   546
          ths |> filter_out (is_likely_tautology ctxt prover orf is_too_meta)
blanchet@49331
   547
              |> map (rpair () o Thm.get_name_hint)
blanchet@49331
   548
              |> Symtab.make
blanchet@49333
   549
        fun do_fact _ (accum as (_, true)) = accum
blanchet@49333
   550
          | do_fact ((_, (_, status)), th) ((parents, upds), false) =
blanchet@49333
   551
            let
blanchet@49333
   552
              val name = Thm.get_name_hint th
blanchet@49333
   553
              val feats = features_of ctxt prover thy status [prop_of th]
blanchet@49333
   554
              val deps = isabelle_dependencies_of all_names th
blanchet@49333
   555
              val upd = (name, parents, feats, deps)
blanchet@49333
   556
            in (([name], upd :: upds), timed_out pass1_learn_timeout_factor) end
blanchet@49333
   557
        val parents = parents_wrt_facts facts fact_graph
blanchet@49333
   558
        val ((_, upds), _) =
blanchet@49333
   559
          ((parents, []), false) |> fold do_fact new_facts |>> apsnd rev
blanchet@49333
   560
        val n = length upds
blanchet@49331
   561
        fun trans {thys, fact_graph} =
blanchet@49331
   562
          let
blanchet@49331
   563
            val mash_INIT_or_ADD =
blanchet@49331
   564
              if Graph.is_empty fact_graph then mash_INIT else mash_ADD
blanchet@49331
   565
            val (upds, fact_graph) =
blanchet@49331
   566
              ([], fact_graph) |> fold (update_fact_graph ctxt) upds
blanchet@49331
   567
          in
blanchet@49331
   568
            (mash_INIT_or_ADD ctxt overlord (rev upds);
blanchet@49331
   569
             {thys = thys |> add_thys_for thy,
blanchet@49331
   570
              fact_graph = fact_graph})
blanchet@49331
   571
          end
blanchet@49333
   572
      in
blanchet@49334
   573
        mash_map trans;
blanchet@49334
   574
        if verbose then
blanchet@49334
   575
          "Processed " ^ string_of_int n ^ " proof" ^ plural_s n ^
blanchet@49334
   576
          (if verbose then
blanchet@49334
   577
             " in " ^ string_from_time (Timer.checkRealTimer timer)
blanchet@49334
   578
           else
blanchet@49334
   579
             "") ^ "."
blanchet@49334
   580
        else
blanchet@49334
   581
          ""
blanchet@49333
   582
      end
blanchet@49323
   583
  end
blanchet@49319
   584
blanchet@49333
   585
fun mash_learn_proof ctxt ({provers, overlord, ...} : params) t used_ths facts =
blanchet@49331
   586
  let
blanchet@49331
   587
    val thy = Proof_Context.theory_of ctxt
blanchet@49333
   588
    val prover = hd provers
blanchet@49333
   589
    val name = ATP_Util.timestamp () ^ serial_string () (* fresh enough *)
blanchet@49333
   590
    val feats = features_of ctxt prover thy General [t]
blanchet@49331
   591
    val deps = used_ths |> map Thm.get_name_hint
blanchet@49331
   592
  in
blanchet@49331
   593
    mash_map (fn {thys, fact_graph} =>
blanchet@49324
   594
        let
blanchet@49333
   595
          val parents = parents_wrt_facts facts fact_graph
blanchet@49331
   596
          val upds = [(name, parents, feats, deps)]
blanchet@49331
   597
          val (upds, fact_graph) =
blanchet@49331
   598
            ([], fact_graph) |> fold (update_fact_graph ctxt) upds
blanchet@49324
   599
        in
blanchet@49331
   600
          mash_ADD ctxt overlord upds;
blanchet@49331
   601
          {thys = thys, fact_graph = fact_graph}
blanchet@49331
   602
        end)
blanchet@49331
   603
  end
blanchet@49264
   604
blanchet@49333
   605
(* The threshold should be large enough so that MaSh doesn't kick in for Auto
blanchet@49333
   606
   Sledgehammer and Try. *)
blanchet@49333
   607
val min_secs_for_learning = 15
blanchet@49333
   608
val short_learn_timeout_factor = 0.2
blanchet@49333
   609
val long_learn_timeout_factor = 4.0
blanchet@49333
   610
blanchet@49333
   611
fun relevant_facts ctxt (params as {fact_filter, timeout, ...}) prover max_facts
blanchet@49313
   612
        ({add, only, ...} : fact_override) hyp_ts concl_t facts =
blanchet@49329
   613
  if not (subset (op =) (the_list fact_filter, fact_filters)) then
blanchet@49329
   614
    error ("Unknown fact filter: " ^ quote (the fact_filter) ^ ".")
blanchet@49329
   615
  else if only then
blanchet@49304
   616
    facts
blanchet@49308
   617
  else if max_facts <= 0 then
blanchet@49303
   618
    []
blanchet@49303
   619
  else
blanchet@49303
   620
    let
blanchet@49333
   621
      val thy = Proof_Context.theory_of ctxt
blanchet@49333
   622
      fun maybe_learn can_suggest =
blanchet@49334
   623
        if Async_Manager.has_running_threads MaShN orelse null facts then
blanchet@49334
   624
          ()
blanchet@49334
   625
        else if Time.toSeconds timeout >= min_secs_for_learning then
blanchet@49334
   626
          let
blanchet@49334
   627
            val factor =
blanchet@49334
   628
              if can_suggest then short_learn_timeout_factor
blanchet@49334
   629
              else long_learn_timeout_factor
blanchet@49334
   630
            val soft_timeout = time_mult factor timeout
blanchet@49334
   631
            val hard_timeout = time_mult 2.0 soft_timeout
blanchet@49334
   632
            val birth_time = Time.now ()
blanchet@49334
   633
            val death_time = Time.+ (birth_time, hard_timeout)
blanchet@49334
   634
            val desc = ("machine learner for Sledgehammer", "")
blanchet@49334
   635
          in
blanchet@49334
   636
            Async_Manager.launch MaShN birth_time death_time desc
blanchet@49334
   637
                (fn () =>
blanchet@49334
   638
                    (true, mash_learn_thy ctxt params thy soft_timeout facts))
blanchet@49334
   639
          end
blanchet@49333
   640
        else
blanchet@49333
   641
          ()
blanchet@49329
   642
      val fact_filter =
blanchet@49329
   643
        case fact_filter of
blanchet@49333
   644
          SOME ff =>
blanchet@49333
   645
          (if ff <> iterN then maybe_learn (mash_can_suggest_facts ()) else ();
blanchet@49333
   646
           ff)
blanchet@49333
   647
        | NONE =>
blanchet@49333
   648
          if mash_can_suggest_facts () then (maybe_learn true; meshN)
blanchet@49333
   649
          else if mash_could_suggest_facts () then (maybe_learn false; iterN)
blanchet@49333
   650
          else iterN
blanchet@49303
   651
      val add_ths = Attrib.eval_thms ctxt add
blanchet@49307
   652
      fun prepend_facts ths accepts =
blanchet@49303
   653
        ((facts |> filter (member Thm.eq_thm_prop ths o snd)) @
blanchet@49307
   654
         (accepts |> filter_out (member Thm.eq_thm_prop ths o snd)))
blanchet@49308
   655
        |> take max_facts
blanchet@49329
   656
      fun iter () =
blanchet@49313
   657
        iterative_relevant_facts ctxt params prover max_facts NONE hyp_ts
blanchet@49313
   658
                                 concl_t facts
blanchet@49329
   659
      fun mash () =
blanchet@49335
   660
        mash_suggest_facts ctxt params prover max_facts hyp_ts concl_t facts
blanchet@49329
   661
      val mess =
blanchet@49335
   662
        [] |> (if fact_filter <> mashN then cons (iter (), []) else I)
blanchet@49329
   663
           |> (if fact_filter <> iterN then cons (mash ()) else I)
blanchet@49303
   664
    in
blanchet@49328
   665
      mesh_facts max_facts mess
blanchet@49303
   666
      |> not (null add_ths) ? prepend_facts add_ths
blanchet@49303
   667
    end
blanchet@49303
   668
blanchet@49334
   669
fun kill_learners () = Async_Manager.kill_threads MaShN "learner"
blanchet@49334
   670
fun running_learners () = Async_Manager.running_threads MaShN "learner"
blanchet@49334
   671
blanchet@49263
   672
end;