(* Title: HOL/TPTP/mash_eval.ML
Author: Jasmin Blanchette, TU Muenchen
Copyright 2012
Evaluate proof suggestions from MaSh (Machine-learning for Sledgehammer).
*)
signature MASH_EVAL =
sig
type params = Sledgehammer_Provers.params
val evaluate_mash_suggestions :
Proof.context -> params -> int * int option -> string option -> string
-> string -> unit
end;
structure MaSh_Eval : MASH_EVAL =
struct
open Sledgehammer_Util
open Sledgehammer_Fact
open Sledgehammer_MePo
open Sledgehammer_MaSh
open Sledgehammer_Provers
open Sledgehammer_Isar
val MePoN = "MePo"
val MaShN = "MaSh"
val MeShN = "MeSh"
val IsarN = "Isar"
fun in_range (from, to) j =
j >= from andalso (to = NONE orelse j <= the to)
fun evaluate_mash_suggestions ctxt params range prob_dir_name sugg_file_name
report_file_name =
let
val report_path = report_file_name |> Path.explode
val _ = File.write report_path ""
fun print s = File.append report_path (s ^ "\n")
val {provers, max_facts, slice, type_enc, lam_trans, timeout, ...} =
default_params ctxt []
val prover = hd provers
val slack_max_facts = generous_max_facts (the max_facts)
val lines = Path.explode sugg_file_name |> File.read_lines
val css = clasimpset_rule_table_of ctxt
val facts = all_facts ctxt true false Symtab.empty [] [] css
val name_tabs = build_name_tables nickname_of_thm facts
fun with_index facts s = (find_index (curry (op =) s) facts + 1, s)
fun index_string (j, s) = s ^ "@" ^ string_of_int j
fun str_of_res label facts ({outcome, run_time, used_facts, ...}
: prover_result) =
let val facts = facts |> map (fn ((name, _), _) => name ()) in
" " ^ label ^ ": " ^
(if is_none outcome then
"Success (" ^ ATP_Util.string_from_time run_time ^ "): " ^
(used_facts |> map (with_index facts o fst)
|> sort (int_ord o pairself fst)
|> map index_string
|> space_implode " ") ^
(if length facts < the max_facts then
" (of " ^ string_of_int (length facts) ^ ")"
else
"")
else
"Failure: " ^
(facts |> take (the max_facts) |> tag_list 1
|> map index_string
|> space_implode " "))
end
fun solve_goal (j, line) =
if in_range range j then
let
val (name, suggs) = extract_suggestions line
val th =
case find_first (fn (_, th) => nickname_of_thm th = name) facts of
SOME (_, th) => th
| NONE => error ("No fact called \"" ^ name ^ "\".")
val goal = goal_of_thm (Proof_Context.theory_of ctxt) th
val (_, hyp_ts, concl_t) = ATP_Util.strip_subgoal ctxt goal 1
val isar_deps = isar_dependencies_of name_tabs th
val facts = facts |> filter (fn (_, th') => thm_ord (th', th) = LESS)
val mepo_facts =
mepo_suggested_facts ctxt params prover slack_max_facts NONE hyp_ts
concl_t facts
|> weight_mepo_facts
val (mash_facts, mash_unks) =
find_mash_suggestions slack_max_facts suggs facts [] []
|>> weight_mash_facts
val mess =
[(mepo_weight, (mepo_facts, [])),
(mash_weight, (mash_facts, mash_unks))]
val mesh_facts =
mesh_facts (Thm.eq_thm_prop o pairself snd) slack_max_facts mess
val isar_facts =
find_suggested_facts (map (rpair 1.0) isar_deps) facts
(* adapted from "mirabelle_sledgehammer.ML" *)
fun set_file_name heading (SOME dir) =
let
val prob_prefix =
"goal_" ^ string_of_int j ^ "__" ^ encode_str name ^ "__" ^
heading
in
Config.put dest_dir dir
#> Config.put problem_prefix (prob_prefix ^ "__")
#> Config.put SMT_Config.debug_files (dir ^ "/" ^ prob_prefix)
end
| set_file_name _ NONE = I
fun prove heading get facts =
let
fun nickify ((_, stature), th) =
((K (encode_str (nickname_of_thm th)), stature), th)
val facts =
facts
|> map (get #> nickify)
|> maybe_instantiate_inducts ctxt hyp_ts concl_t
|> take (the max_facts)
val ctxt = ctxt |> set_file_name heading prob_dir_name
val res as {outcome, ...} =
run_prover_for_mash ctxt params prover facts goal
val ok = if is_none outcome then 1 else 0
in (str_of_res heading facts res, ok) end
val ress =
[fn () => prove MePoN fst mepo_facts,
fn () => prove MaShN fst mash_facts,
fn () => prove MeShN I mesh_facts,
fn () => prove IsarN fst isar_facts]
|> (* Par_List. *) map (fn f => f ())
in
"Goal " ^ string_of_int j ^ ": " ^ name :: map fst ress
|> cat_lines |> print;
map snd ress
end
else
[0, 0, 0, 0]
fun total_of heading ok n =
" " ^ heading ^ ": " ^ string_of_int ok ^ " (" ^
Real.fmt (StringCvt.FIX (SOME 1))
(100.0 * Real.fromInt ok / Real.fromInt n) ^ "%)"
val inst_inducts = Config.get ctxt instantiate_inducts
val options =
["prover = " ^ prover,
"max_facts = " ^ string_of_int (the max_facts),
"slice" |> not slice ? prefix "dont_",
"type_enc = " ^ the_default "smart" type_enc,
"lam_trans = " ^ the_default "smart" lam_trans,
"timeout = " ^ ATP_Util.string_from_time (the_default one_year timeout),
"instantiate_inducts" |> not inst_inducts ? prefix "dont_"]
val _ = print " * * *";
val _ = print ("Options: " ^ commas options);
val oks = Par_List.map solve_goal (tag_list 1 lines)
val n = length oks
val [mepo_ok, mash_ok, mesh_ok, isar_ok] =
map Integer.sum (map_transpose I oks)
in
["Successes (of " ^ string_of_int n ^ " goals)",
total_of MePoN mepo_ok n,
total_of MaShN mash_ok n,
total_of MeShN mesh_ok n,
total_of IsarN isar_ok n]
|> cat_lines |> print
end
end;