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-- |
48285 | 1 |
(* Title: HOL/TPTP/mash_eval.ML |
48234 | 2 |
Author: Jasmin Blanchette, TU Muenchen |
3 |
Copyright 2012 |
|
4 |
||
48285 | 5 |
Evaluate proof suggestions from MaSh (Machine-learning for Sledgehammer). |
48234 | 6 |
*) |
7 |
||
48285 | 8 |
signature MASH_EVAL = |
48234 | 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 | 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 | 14 |
end; |
15 |
||
48285 | 16 |
structure MaSh_Eval : MASH_EVAL = |
48234 | 17 |
struct |
48235 | 18 |
|
48251
6cdcfbddc077
moved most of MaSh exporter code to Sledgehammer
blanchet
parents:
48250
diff
changeset
|
19 |
open Sledgehammer_Filter_MaSh |
48240 | 20 |
|
48235 | 21 |
val unescape_meta = |
22 |
let |
|
23 |
fun un accum [] = String.implode (rev accum) |
|
24 |
| un accum (#"\\" :: d1 :: d2 :: d3 :: cs) = |
|
25 |
(case Int.fromString (String.implode [d1, d2, d3]) of |
|
26 |
SOME n => un (Char.chr n :: accum) cs |
|
27 |
| NONE => "" (* error *)) |
|
28 |
| un _ (#"\\" :: _) = "" (* error *) |
|
29 |
| un accum (c :: cs) = un (c :: accum) cs |
|
30 |
in un [] o String.explode end |
|
31 |
||
32 |
val of_fact_name = unescape_meta |
|
33 |
||
48285 | 34 |
val isaN = "Isa" |
35 |
val iterN = "Iter" |
|
48241 | 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 | 38 |
|
48239
0016290f904c
generate Meng--Paulson facts for evaluation purposes
blanchet
parents:
48236
diff
changeset
|
39 |
val max_relevant_slack = 2 |
48235 | 40 |
|
48285 | 41 |
fun evaluate_mash_suggestions ctxt params thy file_name = |
48235 | 42 |
let |
48245 | 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 | 46 |
val path = file_name |> Path.explode |
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 | 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 | 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 | 54 |
fun find_sugg facts sugg = |
55 |
find_first (fn (_, th) => Thm.get_name_hint th = sugg) facts |
|
48236 | 56 |
fun sugg_facts hyp_ts concl_t facts = |
48235 | 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 | 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 | 62 |
let |
63 |
val fact_eq = (op =) o pairself fst |
|
64 |
fun score_in f fs = |
|
65 |
case find_index (curry fact_eq f) fs of |
|
66 |
~1 => length fs |
|
67 |
| j => j |
|
68 |
fun score_of f = score_in f fs1 + score_in f fs2 |
|
69 |
in |
|
70 |
union fact_eq fs1 fs2 |
|
48236 | 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 | 73 |
end |
74 |
fun with_index facts s = |
|
75 |
(find_index (fn ((s', _), _) => s = s') facts + 1, s) |
|
76 |
fun index_string (j, s) = s ^ "@" ^ string_of_int j |
|
48240 | 77 |
fun str_of_res label facts {outcome, run_time, used_facts, ...} = |
78 |
" " ^ label ^ ": " ^ |
|
79 |
(if is_none outcome then |
|
80 |
"Success (" ^ ATP_Util.string_from_time run_time ^ "): " ^ |
|
48241 | 81 |
(used_facts |> map (with_index facts o fst) |
82 |
|> sort (int_ord o pairself fst) |
|
83 |
|> map index_string |
|
84 |
|> space_implode " ") ^ |
|
85 |
(if length facts < the max_relevant then |
|
86 |
" (of " ^ string_of_int (length facts) ^ ")" |
|
87 |
else |
|
88 |
"") |
|
48240 | 89 |
else |
48241 | 90 |
"Failure: " ^ |
91 |
(facts |> map (fst o fst) |
|
48245 | 92 |
|> take (the max_relevant) |
48241 | 93 |
|> tag_list 1 |
94 |
|> map index_string |
|
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 | 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 | 100 |
val _ = if is_none outcome then ok := !ok + 1 else () |
101 |
in str_of_res heading facts res end |
|
102 |
fun solve_goal j name suggs = |
|
48235 | 103 |
let |
104 |
val name = of_fact_name name |
|
105 |
val th = |
|
106 |
case find_first (fn (_, th) => Thm.get_name_hint th = name) facts of |
|
107 |
SOME (_, th) => th |
|
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 | 111 |
val (_, hyp_ts, concl_t) = ATP_Util.strip_subgoal ctxt goal 1 |
48235 | 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 | 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 | 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 | 126 |
|> cat_lines |> tracing |
48235 | 127 |
end |
128 |
val explode_suggs = space_explode " " #> filter_out (curry (op =) "") |
|
48241 | 129 |
fun do_line (j, line) = |
48235 | 130 |
case space_explode ":" line of |
48241 | 131 |
[goal_name, suggs] => solve_goal j goal_name (explode_suggs suggs) |
48235 | 132 |
| _ => () |
48241 | 133 |
fun total_of heading ok n = |
134 |
" " ^ heading ^ ": " ^ string_of_int (!ok) ^ " (" ^ |
|
135 |
Real.fmt (StringCvt.FIX (SOME 1)) |
|
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 | 138 |
val options = |
48241 | 139 |
[prover_name, string_of_int (the max_relevant) ^ " facts", |
140 |
"slice" |> not slice ? prefix "dont_", the_default "smart" type_enc, |
|
141 |
the_default "smart" lam_trans, ATP_Util.string_from_time timeout, |
|
142 |
"instantiate_inducts" |> not inst_inducts ? prefix "dont_"] |
|
143 |
val n = length lines |
|
144 |
in |
|
145 |
tracing " * * *"; |
|
48245 | 146 |
tracing ("Options: " ^ commas options); |
48241 | 147 |
List.app do_line (tag_list 1 lines); |
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 | 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 | 153 |
|> cat_lines |> tracing |
154 |
end |
|
48235 | 155 |
|
48234 | 156 |
end; |