src/HOL/TPTP/mash_eval.ML
author blanchet
Wed, 18 Jul 2012 08:44:03 +0200
changeset 48285 902ab51dd12a
parent 48284 src/HOL/TPTP/mash_import.ML@a3cb8901d60c
child 48289 6b65f1ad0e4b
permissions -rw-r--r--
renaming
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
48285
902ab51dd12a renaming
blanchet
parents: 48284
diff changeset
     1
(*  Title:      HOL/TPTP/mash_eval.ML
48234
06216c789ac9 moved MaSh into own files
blanchet
parents:
diff changeset
     2
    Author:     Jasmin Blanchette, TU Muenchen
06216c789ac9 moved MaSh into own files
blanchet
parents:
diff changeset
     3
    Copyright   2012
06216c789ac9 moved MaSh into own files
blanchet
parents:
diff changeset
     4
48285
902ab51dd12a renaming
blanchet
parents: 48284
diff changeset
     5
Evaluate proof suggestions from MaSh (Machine-learning for Sledgehammer).
48234
06216c789ac9 moved MaSh into own files
blanchet
parents:
diff changeset
     6
*)
06216c789ac9 moved MaSh into own files
blanchet
parents:
diff changeset
     7
48285
902ab51dd12a renaming
blanchet
parents: 48284
diff changeset
     8
signature MASH_EVAL =
48234
06216c789ac9 moved MaSh into own files
blanchet
parents:
diff changeset
     9
sig
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
    10
  type params = Sledgehammer_Provers.params
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
    11
48285
902ab51dd12a renaming
blanchet
parents: 48284
diff changeset
    12
  val evaluate_mash_suggestions :
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
    13
    Proof.context -> params -> theory -> string -> unit
48234
06216c789ac9 moved MaSh into own files
blanchet
parents:
diff changeset
    14
end;
06216c789ac9 moved MaSh into own files
blanchet
parents:
diff changeset
    15
48285
902ab51dd12a renaming
blanchet
parents: 48284
diff changeset
    16
structure MaSh_Eval : MASH_EVAL =
48234
06216c789ac9 moved MaSh into own files
blanchet
parents:
diff changeset
    17
struct
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    18
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
    19
open Sledgehammer_Filter_MaSh
48240
6a8d18798161 rationalized output
blanchet
parents: 48239
diff changeset
    20
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    21
val unescape_meta =
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    22
  let
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    23
    fun un accum [] = String.implode (rev accum)
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    24
      | un accum (#"\\" :: d1 :: d2 :: d3 :: cs) =
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    25
        (case Int.fromString (String.implode [d1, d2, d3]) of
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    26
           SOME n => un (Char.chr n :: accum) cs
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    27
         | NONE => "" (* error *))
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    28
      | un _ (#"\\" :: _) = "" (* error *)
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    29
      | un accum (c :: cs) = un (c :: accum) cs
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    30
  in un [] o String.explode end
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    31
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    32
val of_fact_name = unescape_meta
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    33
48285
902ab51dd12a renaming
blanchet
parents: 48284
diff changeset
    34
val isaN = "Isa"
902ab51dd12a renaming
blanchet
parents: 48284
diff changeset
    35
val iterN = "Iter"
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    36
val mashN = "MaSh"
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
    37
val iter_mashN = "Iter+MaSh"
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    38
48239
0016290f904c generate Meng--Paulson facts for evaluation purposes
blanchet
parents: 48236
diff changeset
    39
val max_relevant_slack = 2
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    40
48285
902ab51dd12a renaming
blanchet
parents: 48284
diff changeset
    41
fun evaluate_mash_suggestions ctxt params thy file_name =
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    42
  let
48245
854a47677335 generate ATP dependencies
blanchet
parents: 48241
diff changeset
    43
    val {provers, max_relevant, slice, type_enc, lam_trans, timeout, ...} =
48239
0016290f904c generate Meng--Paulson facts for evaluation purposes
blanchet
parents: 48236
diff changeset
    44
      Sledgehammer_Isar.default_params ctxt []
0016290f904c generate Meng--Paulson facts for evaluation purposes
blanchet
parents: 48236
diff changeset
    45
    val prover_name = hd provers
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    46
    val path = file_name |> Path.explode
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    47
    val lines = path |> File.read_lines
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
    48
    val facts = all_non_tautological_facts_of thy
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    49
    val all_names = facts |> map (Thm.get_name_hint o snd)
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
    50
    val iter_ok = Unsynchronized.ref 0
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    51
    val mash_ok = Unsynchronized.ref 0
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
    52
    val iter_mash_ok = Unsynchronized.ref 0
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
    53
    val isa_ok = Unsynchronized.ref 0
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    54
    fun find_sugg facts sugg =
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    55
      find_first (fn (_, th) => Thm.get_name_hint th = sugg) facts
48236
e174ecc4f5a4 instantiate induction rules
blanchet
parents: 48235
diff changeset
    56
    fun sugg_facts hyp_ts concl_t facts =
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    57
      map_filter (find_sugg facts o of_fact_name)
48239
0016290f904c generate Meng--Paulson facts for evaluation purposes
blanchet
parents: 48236
diff changeset
    58
      #> take (max_relevant_slack * the max_relevant)
48250
1065c307fafe further ML structure split to permit finer-grained loading/reordering (problem to solve: MaSh needs most of Sledgehammer)
blanchet
parents: 48245
diff changeset
    59
      #> Sledgehammer_Fact.maybe_instantiate_inducts ctxt hyp_ts concl_t
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    60
      #> map (apfst (apfst (fn name => name ())))
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
    61
    fun iter_mash_facts fs1 fs2 =
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    62
      let
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    63
        val fact_eq = (op =) o pairself fst
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    64
        fun score_in f fs =
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    65
          case find_index (curry fact_eq f) fs of
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    66
            ~1 => length fs
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    67
          | j => j
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    68
        fun score_of f = score_in f fs1 + score_in f fs2
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    69
      in
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    70
        union fact_eq fs1 fs2
48236
e174ecc4f5a4 instantiate induction rules
blanchet
parents: 48235
diff changeset
    71
        |> map (`score_of) |> sort (int_ord o pairself fst) |> map snd
48239
0016290f904c generate Meng--Paulson facts for evaluation purposes
blanchet
parents: 48236
diff changeset
    72
        |> take (the max_relevant)
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    73
      end
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    74
    fun with_index facts s =
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    75
      (find_index (fn ((s', _), _) => s = s') facts + 1, s)
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    76
    fun index_string (j, s) = s ^ "@" ^ string_of_int j
48240
6a8d18798161 rationalized output
blanchet
parents: 48239
diff changeset
    77
    fun str_of_res label facts {outcome, run_time, used_facts, ...} =
6a8d18798161 rationalized output
blanchet
parents: 48239
diff changeset
    78
      "  " ^ label ^ ": " ^
6a8d18798161 rationalized output
blanchet
parents: 48239
diff changeset
    79
      (if is_none outcome then
6a8d18798161 rationalized output
blanchet
parents: 48239
diff changeset
    80
         "Success (" ^ ATP_Util.string_from_time run_time ^ "): " ^
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    81
         (used_facts |> map (with_index facts o fst)
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    82
                     |> sort (int_ord o pairself fst)
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    83
                     |> map index_string
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    84
                     |> space_implode " ") ^
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    85
         (if length facts < the max_relevant then
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    86
            " (of " ^ string_of_int (length facts) ^ ")"
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    87
          else
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    88
            "")
48240
6a8d18798161 rationalized output
blanchet
parents: 48239
diff changeset
    89
       else
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    90
         "Failure: " ^
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    91
         (facts |> map (fst o fst)
48245
854a47677335 generate ATP dependencies
blanchet
parents: 48241
diff changeset
    92
                |> take (the max_relevant)
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    93
                |> tag_list 1
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    94
                |> map index_string
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
    95
                |> space_implode " "))
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
    96
    fun prove ok heading facts goal =
48239
0016290f904c generate Meng--Paulson facts for evaluation purposes
blanchet
parents: 48236
diff changeset
    97
      let
48245
854a47677335 generate ATP dependencies
blanchet
parents: 48241
diff changeset
    98
        val facts = facts |> take (the max_relevant)
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
    99
        val res as {outcome, ...} = run_prover ctxt params facts goal
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   100
        val _ = if is_none outcome then ok := !ok + 1 else ()
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   101
      in str_of_res heading facts res end
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   102
    fun solve_goal j name suggs =
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
   103
      let
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
   104
        val name = of_fact_name name
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
   105
        val th =
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
   106
          case find_first (fn (_, th) => Thm.get_name_hint th = name) facts of
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
   107
            SOME (_, th) => th
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
   108
          | NONE => error ("No fact called \"" ^ name ^ "\"")
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   109
        val isa_deps = isabelle_dependencies_of all_names th
48250
1065c307fafe further ML structure split to permit finer-grained loading/reordering (problem to solve: MaSh needs most of Sledgehammer)
blanchet
parents: 48245
diff changeset
   110
        val goal = goal_of_thm thy th
48245
854a47677335 generate ATP dependencies
blanchet
parents: 48241
diff changeset
   111
        val (_, hyp_ts, concl_t) = ATP_Util.strip_subgoal ctxt goal 1
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
   112
        val facts = facts |> filter (fn (_, th') => thm_ord (th', th) = LESS)
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   113
        val isa_facts = sugg_facts hyp_ts concl_t facts isa_deps
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   114
        val iter_facts =
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   115
          iter_facts ctxt params (max_relevant_slack * the max_relevant) goal
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   116
                     facts
48236
e174ecc4f5a4 instantiate induction rules
blanchet
parents: 48235
diff changeset
   117
        val mash_facts = sugg_facts hyp_ts concl_t facts suggs
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   118
        val iter_mash_facts = iter_mash_facts iter_facts mash_facts
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   119
        val iter_s = prove iter_ok iterN iter_facts goal
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   120
        val mash_s = prove mash_ok mashN mash_facts goal
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   121
        val iter_mash_s = prove iter_mash_ok iter_mashN iter_mash_facts goal
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   122
        val isa_s = prove isa_ok isaN isa_facts goal
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
   123
      in
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   124
        ["Goal " ^ string_of_int j ^ ": " ^ name, iter_s, mash_s, iter_mash_s,
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   125
         isa_s]
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   126
        |> cat_lines |> tracing
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
   127
      end
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
   128
    val explode_suggs = space_explode " " #> filter_out (curry (op =) "")
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   129
    fun do_line (j, line) =
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
   130
      case space_explode ":" line of
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   131
        [goal_name, suggs] => solve_goal j goal_name (explode_suggs suggs)
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
   132
      | _ => ()
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   133
    fun total_of heading ok n =
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   134
      " " ^ heading ^ ": " ^ string_of_int (!ok) ^ " (" ^
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   135
      Real.fmt (StringCvt.FIX (SOME 1))
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   136
               (100.0 * Real.fromInt (!ok) / Real.fromInt n) ^ "%)"
48250
1065c307fafe further ML structure split to permit finer-grained loading/reordering (problem to solve: MaSh needs most of Sledgehammer)
blanchet
parents: 48245
diff changeset
   137
    val inst_inducts = Config.get ctxt Sledgehammer_Fact.instantiate_inducts
48245
854a47677335 generate ATP dependencies
blanchet
parents: 48241
diff changeset
   138
    val options =
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   139
      [prover_name, string_of_int (the max_relevant) ^ " facts",
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   140
       "slice" |> not slice ? prefix "dont_", the_default "smart" type_enc,
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   141
       the_default "smart" lam_trans, ATP_Util.string_from_time timeout,
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   142
       "instantiate_inducts" |> not inst_inducts ? prefix "dont_"]
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   143
    val n = length lines
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   144
  in
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   145
    tracing " * * *";
48245
854a47677335 generate ATP dependencies
blanchet
parents: 48241
diff changeset
   146
    tracing ("Options: " ^ commas options);
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   147
    List.app do_line (tag_list 1 lines);
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   148
    ["Successes (of " ^ string_of_int n ^ " goals)",
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   149
     total_of iterN iter_ok n,
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   150
     total_of mashN mash_ok n,
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   151
     total_of iter_mashN iter_mash_ok n,
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
   152
     total_of isaN isa_ok n]
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   153
    |> cat_lines |> tracing
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   154
  end
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
   155
48234
06216c789ac9 moved MaSh into own files
blanchet
parents:
diff changeset
   156
end;