src/HOL/TPTP/mash_eval.ML
author blanchet
Sun, 04 Oct 2015 17:41:52 +0200
changeset 61321 c982a4cc8dc4
parent 60641 f6e8293747fd
child 63692 1bc4bc2c9fd1
permissions -rw-r--r--
sped up MaSh nickname generation
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
55201
1ee776da8da7 renamed ML file
blanchet
parents: 55198
diff changeset
    10
  type params = Sledgehammer_Prover.params
48251
6cdcfbddc077 moved most of MaSh exporter code to Sledgehammer
blanchet
parents: 48250
diff changeset
    11
57456
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
    12
  val evaluate_mash_suggestions : Proof.context -> params -> int * int option -> string option ->
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
    13
    string list -> 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
50557
31313171deb5 thread no timeout properly
blanchet
parents: 50556
diff changeset
    19
open Sledgehammer_Util
48315
82d6e46c673f fixed order of accessibles + other tweaks to MaSh
blanchet
parents: 48313
diff changeset
    20
open Sledgehammer_Fact
50557
31313171deb5 thread no timeout properly
blanchet
parents: 50556
diff changeset
    21
open Sledgehammer_MePo
48381
1b7d798460bb renamed ML structures
blanchet
parents: 48379
diff changeset
    22
open Sledgehammer_MaSh
55201
1ee776da8da7 renamed ML file
blanchet
parents: 55198
diff changeset
    23
open Sledgehammer_Prover
55212
blanchet
parents: 55205
diff changeset
    24
open Sledgehammer_Prover_ATP
55198
7a538e58b64e tuned ML file name
blanchet
parents: 54816
diff changeset
    25
open Sledgehammer_Commands
57431
02c408aed5ee killed Python version of MaSh, now that the SML version works adequately
blanchet
parents: 57406
diff changeset
    26
open MaSh_Export
48240
6a8d18798161 rationalized output
blanchet
parents: 48239
diff changeset
    27
53980
7e6a82c593f4 made SML/NJ happy (NB: toplevel ML environment is unmanaged);
wenzelm
parents: 53139
diff changeset
    28
val prefix = Library.prefix
7e6a82c593f4 made SML/NJ happy (NB: toplevel ML environment is unmanaged);
wenzelm
parents: 53139
diff changeset
    29
57456
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
    30
fun evaluate_mash_suggestions ctxt params range prob_dir_name file_names report_file_name =
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
    31
  let
55205
8450622db0c5 refactor large ML file
blanchet
parents: 55201
diff changeset
    32
    val thy = Proof_Context.theory_of ctxt
51028
7327a847f0c7 more robustness w.r.t. 0
blanchet
parents: 51015
diff changeset
    33
    val zeros = [0, 0, 0, 0, 0, 0]
50448
0a740d127187 have MaSh evaluator keep all raw problem/solution files in a directory
blanchet
parents: 50442
diff changeset
    34
    val report_path = report_file_name |> Path.explode
0a740d127187 have MaSh evaluator keep all raw problem/solution files in a directory
blanchet
parents: 50442
diff changeset
    35
    val _ = File.write report_path ""
57432
blanchet
parents: 57431
diff changeset
    36
50589
42f3630a6a0f no need for tracing
blanchet
parents: 50587
diff changeset
    37
    fun print s = File.append report_path (s ^ "\n")
57432
blanchet
parents: 57431
diff changeset
    38
55205
8450622db0c5 refactor large ML file
blanchet
parents: 55201
diff changeset
    39
    val {provers, max_facts, slice, type_enc, lam_trans, timeout, ...} = default_params thy []
48318
325c8fd0d762 more consolidation of MaSh code
blanchet
parents: 48316
diff changeset
    40
    val prover = hd provers
57108
dc0b4f50e288 more generous max number of suggestions, for more safety
blanchet
parents: 55212
diff changeset
    41
    val max_suggs = generous_max_suggestions (the max_facts)
57456
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
    42
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
    43
    val method_of_file_name =
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
    44
      perhaps (try (unsuffix "_suggestions")) o List.last o space_explode "/"
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
    45
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
    46
    val methods = "isar" :: map method_of_file_name file_names
50964
2a990baa09af use precomputed MaSh/MePo data whenever available
blanchet
parents: 50953
diff changeset
    47
    val lines_of = Path.explode #> try File.read_lines #> these
57456
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
    48
    val liness0 = map lines_of file_names
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
    49
    val num_lines = fold (Integer.max o length) liness0 0
57432
blanchet
parents: 57431
diff changeset
    50
50964
2a990baa09af use precomputed MaSh/MePo data whenever available
blanchet
parents: 50953
diff changeset
    51
    fun pad lines = lines @ replicate (num_lines - length lines) ""
57432
blanchet
parents: 57431
diff changeset
    52
58205
369513534627 added missing 'transpose'
blanchet
parents: 58061
diff changeset
    53
    val liness' = Ctr_Sugar_Util.transpose (map pad liness0)
57456
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
    54
50557
31313171deb5 thread no timeout properly
blanchet
parents: 50556
diff changeset
    55
    val css = clasimpset_rule_table_of ctxt
58922
1f500b18c4c6 proper Keyword.keywords (cf. 82a71046dce8);
wenzelm
parents: 58205
diff changeset
    56
    val facts = all_facts ctxt true false Keyword.empty_keywords [] [] css
61321
c982a4cc8dc4 sped up MaSh nickname generation
blanchet
parents: 60641
diff changeset
    57
    val name_tabs = build_name_tables nickname_of_thm facts
57432
blanchet
parents: 57431
diff changeset
    58
48289
6b65f1ad0e4b systematize lazy names in relevance filter
blanchet
parents: 48285
diff changeset
    59
    fun with_index facts s = (find_index (curry (op =) s) facts + 1, s)
50952
3fd83a0cc4bd evaluate more cases (cf. paper)
blanchet
parents: 50826
diff changeset
    60
    fun index_str (j, s) = s ^ "@" ^ string_of_int j
50953
c4c746bbf836 provide a means to skip a method
blanchet
parents: 50952
diff changeset
    61
    val str_of_method = enclose "  " ": "
57432
blanchet
parents: 57431
diff changeset
    62
blanchet
parents: 57431
diff changeset
    63
    fun str_of_result method facts ({outcome, run_time, used_facts, ...} : prover_result) =
51004
5f2788c38127 distinguish raw and non-raw facts, using raw for 10 000s of facts and non-raw after selection of some hundreds
blanchet
parents: 50967
diff changeset
    64
      let val facts = facts |> map (fst o fst) in
50953
c4c746bbf836 provide a means to skip a method
blanchet
parents: 50952
diff changeset
    65
        str_of_method method ^
48289
6b65f1ad0e4b systematize lazy names in relevance filter
blanchet
parents: 48285
diff changeset
    66
        (if is_none outcome then
52031
9a9238342963 tuning -- renamed '_from_' to '_of_' in Sledgehammer
blanchet
parents: 51199
diff changeset
    67
           "Success (" ^ ATP_Util.string_of_time run_time ^ "): " ^
48289
6b65f1ad0e4b systematize lazy names in relevance filter
blanchet
parents: 48285
diff changeset
    68
           (used_facts |> map (with_index facts o fst)
59058
a78612c67ec0 renamed "pairself" to "apply2", in accordance to @{apply 2};
wenzelm
parents: 58922
diff changeset
    69
                       |> sort (int_ord o apply2 fst)
50952
3fd83a0cc4bd evaluate more cases (cf. paper)
blanchet
parents: 50826
diff changeset
    70
                       |> map index_str
48289
6b65f1ad0e4b systematize lazy names in relevance filter
blanchet
parents: 48285
diff changeset
    71
                       |> space_implode " ") ^
48293
914ca0827804 renamed Sledgehammer options
blanchet
parents: 48292
diff changeset
    72
           (if length facts < the max_facts then
48289
6b65f1ad0e4b systematize lazy names in relevance filter
blanchet
parents: 48285
diff changeset
    73
              " (of " ^ string_of_int (length facts) ^ ")"
6b65f1ad0e4b systematize lazy names in relevance filter
blanchet
parents: 48285
diff changeset
    74
            else
6b65f1ad0e4b systematize lazy names in relevance filter
blanchet
parents: 48285
diff changeset
    75
              "")
6b65f1ad0e4b systematize lazy names in relevance filter
blanchet
parents: 48285
diff changeset
    76
         else
6b65f1ad0e4b systematize lazy names in relevance filter
blanchet
parents: 48285
diff changeset
    77
           "Failure: " ^
48293
914ca0827804 renamed Sledgehammer options
blanchet
parents: 48292
diff changeset
    78
           (facts |> take (the max_facts) |> tag_list 1
50952
3fd83a0cc4bd evaluate more cases (cf. paper)
blanchet
parents: 50826
diff changeset
    79
                  |> map index_str
48289
6b65f1ad0e4b systematize lazy names in relevance filter
blanchet
parents: 48285
diff changeset
    80
                  |> space_implode " "))
6b65f1ad0e4b systematize lazy names in relevance filter
blanchet
parents: 48285
diff changeset
    81
      end
57432
blanchet
parents: 57431
diff changeset
    82
57456
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
    83
    fun solve_goal (j, lines) =
58205
369513534627 added missing 'transpose'
blanchet
parents: 58061
diff changeset
    84
      if in_range range j andalso exists (curry (op <>) "") lines then
50559
89c0d2f13cca MaSh exporter can now export subsets of the facts, as consecutive ranges
blanchet
parents: 50557
diff changeset
    85
        let
57125
2f620ef839ee added another way of invoking Python code, for experiments
blanchet
parents: 57108
diff changeset
    86
          val get_suggs = extract_suggestions ##> (take max_suggs #> map fst)
57456
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
    87
          val (names, suggss0) = split_list (map get_suggs lines)
58205
369513534627 added missing 'transpose'
blanchet
parents: 58061
diff changeset
    88
          val name =
369513534627 added missing 'transpose'
blanchet
parents: 58061
diff changeset
    89
            (case names |> filter (curry (op <>) "") |> distinct (op =) of
369513534627 added missing 'transpose'
blanchet
parents: 58061
diff changeset
    90
              [name] => name
369513534627 added missing 'transpose'
blanchet
parents: 58061
diff changeset
    91
            | names => error ("Input files out of sync: facts " ^ commas (map quote names) ^ "."))
50559
89c0d2f13cca MaSh exporter can now export subsets of the facts, as consecutive ranges
blanchet
parents: 50557
diff changeset
    92
          val th =
61321
c982a4cc8dc4 sped up MaSh nickname generation
blanchet
parents: 60641
diff changeset
    93
            case find_first (fn (_, th) => nickname_of_thm th = name) facts of
50559
89c0d2f13cca MaSh exporter can now export subsets of the facts, as consecutive ranges
blanchet
parents: 50557
diff changeset
    94
              SOME (_, th) => th
89c0d2f13cca MaSh exporter can now export subsets of the facts, as consecutive ranges
blanchet
parents: 50557
diff changeset
    95
            | NONE => error ("No fact called \"" ^ name ^ "\".")
89c0d2f13cca MaSh exporter can now export subsets of the facts, as consecutive ranges
blanchet
parents: 50557
diff changeset
    96
          val goal = goal_of_thm (Proof_Context.theory_of ctxt) th
52196
2281f33e8da6 redid rac7830871177 to avoid duplicate fixed variable (e.g. lemma "P (a::nat)" proof - have "!!a::int. Q a" sledgehammer [e])
blanchet
parents: 52125
diff changeset
    97
          val (_, hyp_ts, concl_t) = ATP_Util.strip_subgoal goal 1 ctxt
61321
c982a4cc8dc4 sped up MaSh nickname generation
blanchet
parents: 60641
diff changeset
    98
          val isar_deps = these (isar_dependencies_of name_tabs th)
57456
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
    99
          val suggss = isar_deps :: suggss0
57406
e844dcf57deb killed dead code
blanchet
parents: 57306
diff changeset
   100
          val facts = facts |> filter (fn (_, th') => thm_less (th', th))
57432
blanchet
parents: 57431
diff changeset
   101
50559
89c0d2f13cca MaSh exporter can now export subsets of the facts, as consecutive ranges
blanchet
parents: 50557
diff changeset
   102
          (* adapted from "mirabelle_sledgehammer.ML" *)
50953
c4c746bbf836 provide a means to skip a method
blanchet
parents: 50952
diff changeset
   103
          fun set_file_name method (SOME dir) =
50559
89c0d2f13cca MaSh exporter can now export subsets of the facts, as consecutive ranges
blanchet
parents: 50557
diff changeset
   104
              let
57432
blanchet
parents: 57431
diff changeset
   105
                val prob_prefix = "goal_" ^ string_of_int j ^ "__" ^ encode_str name ^ "__" ^ method
50559
89c0d2f13cca MaSh exporter can now export subsets of the facts, as consecutive ranges
blanchet
parents: 50557
diff changeset
   106
              in
55212
blanchet
parents: 55205
diff changeset
   107
                Config.put atp_dest_dir dir
blanchet
parents: 55205
diff changeset
   108
                #> Config.put atp_problem_prefix (prob_prefix ^ "__")
58061
3d060f43accb renamed new SMT module from 'SMT2' to 'SMT'
blanchet
parents: 57456
diff changeset
   109
                #> Config.put SMT_Config.debug_files (dir ^ "/" ^ prob_prefix)
50559
89c0d2f13cca MaSh exporter can now export subsets of the facts, as consecutive ranges
blanchet
parents: 50557
diff changeset
   110
              end
89c0d2f13cca MaSh exporter can now export subsets of the facts, as consecutive ranges
blanchet
parents: 50557
diff changeset
   111
            | set_file_name _ NONE = I
57432
blanchet
parents: 57431
diff changeset
   112
53139
07a6e11f1631 cleanup old duplicated functionality
blanchet
parents: 52196
diff changeset
   113
          fun prove method suggs =
57456
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
   114
            if null facts then
50953
c4c746bbf836 provide a means to skip a method
blanchet
parents: 50952
diff changeset
   115
              (str_of_method method ^ "Skipped", 0)
50952
3fd83a0cc4bd evaluate more cases (cf. paper)
blanchet
parents: 50826
diff changeset
   116
            else
3fd83a0cc4bd evaluate more cases (cf. paper)
blanchet
parents: 50826
diff changeset
   117
              let
3fd83a0cc4bd evaluate more cases (cf. paper)
blanchet
parents: 50826
diff changeset
   118
                fun nickify ((_, stature), th) =
61321
c982a4cc8dc4 sped up MaSh nickname generation
blanchet
parents: 60641
diff changeset
   119
                  ((K (encode_str (nickname_of_thm th)), stature), th)
57432
blanchet
parents: 57431
diff changeset
   120
50952
3fd83a0cc4bd evaluate more cases (cf. paper)
blanchet
parents: 50826
diff changeset
   121
                val facts =
53139
07a6e11f1631 cleanup old duplicated functionality
blanchet
parents: 52196
diff changeset
   122
                  suggs
07a6e11f1631 cleanup old duplicated functionality
blanchet
parents: 52196
diff changeset
   123
                  |> find_suggested_facts ctxt facts
07a6e11f1631 cleanup old duplicated functionality
blanchet
parents: 52196
diff changeset
   124
                  |> map (fact_of_raw_fact #> nickify)
50952
3fd83a0cc4bd evaluate more cases (cf. paper)
blanchet
parents: 50826
diff changeset
   125
                  |> maybe_instantiate_inducts ctxt hyp_ts concl_t
3fd83a0cc4bd evaluate more cases (cf. paper)
blanchet
parents: 50826
diff changeset
   126
                  |> take (the max_facts)
51004
5f2788c38127 distinguish raw and non-raw facts, using raw for 10 000s of facts and non-raw after selection of some hundreds
blanchet
parents: 50967
diff changeset
   127
                  |> map fact_of_raw_fact
50953
c4c746bbf836 provide a means to skip a method
blanchet
parents: 50952
diff changeset
   128
                val ctxt = ctxt |> set_file_name method prob_dir_name
57434
6ea8b8592787 use SMT2
blanchet
parents: 57432
diff changeset
   129
                val res as {outcome, ...} = run_prover_for_mash ctxt params prover name facts goal
50952
3fd83a0cc4bd evaluate more cases (cf. paper)
blanchet
parents: 50826
diff changeset
   130
                val ok = if is_none outcome then 1 else 0
57432
blanchet
parents: 57431
diff changeset
   131
              in
blanchet
parents: 57431
diff changeset
   132
                (str_of_result method facts res, ok)
blanchet
parents: 57431
diff changeset
   133
              end
blanchet
parents: 57431
diff changeset
   134
57456
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
   135
          val ress = map2 prove methods suggss
50559
89c0d2f13cca MaSh exporter can now export subsets of the facts, as consecutive ranges
blanchet
parents: 50557
diff changeset
   136
        in
50591
c5e0073558f3 avoid references altogether
blanchet
parents: 50589
diff changeset
   137
          "Goal " ^ string_of_int j ^ ": " ^ name :: map fst ress
50587
bd6582be1562 synchronize access to shared reference and print proper total
blanchet
parents: 50563
diff changeset
   138
          |> cat_lines |> print;
50591
c5e0073558f3 avoid references altogether
blanchet
parents: 50589
diff changeset
   139
          map snd ress
50559
89c0d2f13cca MaSh exporter can now export subsets of the facts, as consecutive ranges
blanchet
parents: 50557
diff changeset
   140
        end
89c0d2f13cca MaSh exporter can now export subsets of the facts, as consecutive ranges
blanchet
parents: 50557
diff changeset
   141
      else
51028
7327a847f0c7 more robustness w.r.t. 0
blanchet
parents: 51015
diff changeset
   142
        zeros
57432
blanchet
parents: 57431
diff changeset
   143
50557
31313171deb5 thread no timeout properly
blanchet
parents: 50556
diff changeset
   144
    val inst_inducts = Config.get ctxt instantiate_inducts
48245
854a47677335 generate ATP dependencies
blanchet
parents: 48241
diff changeset
   145
    val options =
50767
26ad3da13f47 tuned output
blanchet
parents: 50754
diff changeset
   146
      ["prover = " ^ prover,
26ad3da13f47 tuned output
blanchet
parents: 50754
diff changeset
   147
       "max_facts = " ^ string_of_int (the max_facts),
26ad3da13f47 tuned output
blanchet
parents: 50754
diff changeset
   148
       "slice" |> not slice ? prefix "dont_",
26ad3da13f47 tuned output
blanchet
parents: 50754
diff changeset
   149
       "type_enc = " ^ the_default "smart" type_enc,
26ad3da13f47 tuned output
blanchet
parents: 50754
diff changeset
   150
       "lam_trans = " ^ the_default "smart" lam_trans,
54816
10d48c2a3e32 made timeouts in Sledgehammer not be 'option's -- simplified lots of code
blanchet
parents: 54141
diff changeset
   151
       "timeout = " ^ ATP_Util.string_of_time timeout,
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   152
       "instantiate_inducts" |> not inst_inducts ? prefix "dont_"]
50587
bd6582be1562 synchronize access to shared reference and print proper total
blanchet
parents: 50563
diff changeset
   153
    val _ = print " * * *";
bd6582be1562 synchronize access to shared reference and print proper total
blanchet
parents: 50563
diff changeset
   154
    val _ = print ("Options: " ^ commas options);
58205
369513534627 added missing 'transpose'
blanchet
parents: 58061
diff changeset
   155
    val oks = Par_List.map solve_goal (tag_list 1 liness')
50620
07e08250a880 fixed total
blanchet
parents: 50591
diff changeset
   156
    val n = length oks
57456
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
   157
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
   158
    fun total_of method ok =
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
   159
      str_of_method method ^ string_of_int ok ^ " (" ^ Real.fmt (StringCvt.FIX (SOME 1))
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
   160
        (100.0 * Real.fromInt ok / Real.fromInt (Int.max (1, n))) ^ "%)"
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
   161
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
   162
    val oks' = if n = 0 then zeros else map Integer.sum (map_transpose I oks)
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   163
  in
57456
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
   164
    "Successes (of " ^ string_of_int n ^ " goals)" ::
eb5515784992 clean up MaSh evaluation driver
blanchet
parents: 57434
diff changeset
   165
    map2 total_of methods oks'
50437
9762fe72936e store evaluation output in a file
blanchet
parents: 50436
diff changeset
   166
    |> cat_lines |> print
48241
688f1172d523 nicer output
blanchet
parents: 48240
diff changeset
   167
  end
48235
40655464a93b MaSh evaluation driver
blanchet
parents: 48234
diff changeset
   168
48234
06216c789ac9 moved MaSh into own files
blanchet
parents:
diff changeset
   169
end;