[clangd] Use Decision Forest to score code completions.

By default clangd will score a code completion item using heuristics model.

Scoring can be done by Decision Forest model by passing `--ranking_model=decision_forest` to
clangd.

Features omitted from the model:
- `NameMatch` is excluded because the final score must be multiplicative in `NameMatch` to allow rescoring by the editor.
- `NeedsFixIts` is excluded because the generating dataset that needs 'fixits' is non-trivial.

There are multiple ways (heuristics) to combine the above two features with the prediction of the DF:
- `NeedsFixIts` is used as is with a penalty of `0.5`.

Various alternatives of combining NameMatch `N` and Decision forest Prediction `P`
- N * scale(P, 0, 1): Linearly scale the output of model to range [0, 1]
- N * a^P:
  - More natural: Prediction of each Decision Tree can be considered as a multiplicative boost (like NameMatch)
  - Ordering is independent of the absolute value of P. Order of two items is proportional to `a^{difference in model prediction score}`. Higher `a` gives higher weightage to model output as compared to NameMatch score.

Baseline MRR = 0.619
MRR for various combinations:
N * P = 0.6346, advantage%=2.5768
N * 1.1^P = 0.6600, advantage%=6.6853
N * **1.2**^P = 0.6669, advantage%=**7.8005**
N * **1.3**^P = 0.6668, advantage%=**7.7795**
N * **1.4**^P = 0.6659, advantage%=**7.6270**
N * 1.5^P = 0.6646, advantage%=7.4200
N * 1.6^P = 0.6636, advantage%=7.2671
N * 1.7^P = 0.6629, advantage%=7.1450
N * 2^P = 0.6612, advantage%=6.8673
N * 2.5^P = 0.6598, advantage%=6.6491
N * 3^P = 0.6590, advantage%=6.5242
N * scaled[0, 1] = 0.6465, advantage%=4.5054

Differential Revision: https://reviews.llvm.org/D88281
This commit is contained in:
Utkarsh Saxena 2020-09-22 07:56:08 +02:00
parent 76753a597b
commit a8b55b6939
6 changed files with 145 additions and 18 deletions

View File

@ -1625,6 +1625,43 @@ private:
return Filter->match(C.Name);
}
CodeCompletion::Scores
evaluateCompletion(const SymbolQualitySignals &Quality,
const SymbolRelevanceSignals &Relevance) {
using RM = CodeCompleteOptions::CodeCompletionRankingModel;
CodeCompletion::Scores Scores;
switch (Opts.RankingModel) {
case RM::Heuristics:
Scores.Quality = Quality.evaluate();
Scores.Relevance = Relevance.evaluate();
Scores.Total =
evaluateSymbolAndRelevance(Scores.Quality, Scores.Relevance);
// NameMatch is in fact a multiplier on total score, so rescoring is
// sound.
Scores.ExcludingName = Relevance.NameMatch
? Scores.Total / Relevance.NameMatch
: Scores.Quality;
return Scores;
case RM::DecisionForest:
Scores.Quality = 0;
Scores.Relevance = 0;
// Exponentiating DecisionForest prediction makes the score of each tree a
// multiplciative boost (like NameMatch). This allows us to weigh the
// prediciton score and NameMatch appropriately.
Scores.ExcludingName = pow(Opts.DecisionForestBase,
evaluateDecisionForest(Quality, Relevance));
// NeedsFixIts is not part of the DecisionForest as generating training
// data that needs fixits is not-feasible.
if (Relevance.NeedsFixIts)
Scores.ExcludingName *= 0.5;
// NameMatch should be a multiplier on total score to support rescoring.
Scores.Total = Relevance.NameMatch * Scores.ExcludingName;
return Scores;
}
llvm_unreachable("Unhandled CodeCompletion ranking model.");
}
// Scores a candidate and adds it to the TopN structure.
void addCandidate(TopN<ScoredBundle, ScoredBundleGreater> &Candidates,
CompletionCandidate::Bundle Bundle) {
@ -1632,6 +1669,7 @@ private:
SymbolRelevanceSignals Relevance;
Relevance.Context = CCContextKind;
Relevance.Name = Bundle.front().Name;
Relevance.FilterLength = HeuristicPrefix.Name.size();
Relevance.Query = SymbolRelevanceSignals::CodeComplete;
Relevance.FileProximityMatch = FileProximity.getPointer();
if (ScopeProximity)
@ -1680,15 +1718,7 @@ private:
}
}
CodeCompletion::Scores Scores;
Scores.Quality = Quality.evaluate();
Scores.Relevance = Relevance.evaluate();
Scores.Total = evaluateSymbolAndRelevance(Scores.Quality, Scores.Relevance);
// NameMatch is in fact a multiplier on total score, so rescoring is sound.
Scores.ExcludingName = Relevance.NameMatch
? Scores.Total / Relevance.NameMatch
: Scores.Quality;
CodeCompletion::Scores Scores = evaluateCompletion(Quality, Relevance);
if (Opts.RecordCCResult)
Opts.RecordCCResult(toCodeCompletion(Bundle), Quality, Relevance,
Scores.Total);

View File

@ -147,6 +147,22 @@ struct CodeCompleteOptions {
std::function<void(const CodeCompletion &, const SymbolQualitySignals &,
const SymbolRelevanceSignals &, float Score)>
RecordCCResult;
/// Model to use for ranking code completion candidates.
enum CodeCompletionRankingModel {
Heuristics,
DecisionForest,
} RankingModel = Heuristics;
/// Weight for combining NameMatch and Prediction of DecisionForest.
/// CompletionScore is NameMatch * pow(Base, Prediction).
/// The optimal value of Base largely depends on the semantics of the model
/// and prediction score (e.g. algorithm used during training, number of
/// trees, etc.). Usually if the range of Prediciton is [-20, 20] then a Base
/// in [1.2, 1.7] works fine.
/// Semantics: E.g. the completion score reduces by 50% if the Prediciton
/// score is reduced by 2.6 points for Base = 1.3.
float DecisionForestBase = 1.3f;
};
// Semi-structured representation of a code-complete suggestion for our C++ API.

View File

@ -8,6 +8,7 @@
#include "Quality.h"
#include "AST.h"
#include "CompletionModel.h"
#include "FileDistance.h"
#include "SourceCode.h"
#include "URI.h"
@ -486,6 +487,34 @@ float evaluateSymbolAndRelevance(float SymbolQuality, float SymbolRelevance) {
return SymbolQuality * SymbolRelevance;
}
float evaluateDecisionForest(const SymbolQualitySignals &Quality,
const SymbolRelevanceSignals &Relevance) {
Example E;
E.setIsDeprecated(Quality.Deprecated);
E.setIsReservedName(Quality.ReservedName);
E.setIsImplementationDetail(Quality.ImplementationDetail);
E.setNumReferences(Quality.References);
E.setSymbolCategory(Quality.Category);
SymbolRelevanceSignals::DerivedSignals Derived =
Relevance.calculateDerivedSignals();
E.setIsNameInContext(Derived.NameMatchesContext);
E.setIsForbidden(Relevance.Forbidden);
E.setIsInBaseClass(Relevance.InBaseClass);
E.setFileProximityDistance(Derived.FileProximityDistance);
E.setSemaFileProximityScore(Relevance.SemaFileProximityScore);
E.setSymbolScopeDistance(Derived.ScopeProximityDistance);
E.setSemaSaysInScope(Relevance.SemaSaysInScope);
E.setScope(Relevance.Scope);
E.setContextKind(Relevance.Context);
E.setIsInstanceMember(Relevance.IsInstanceMember);
E.setHadContextType(Relevance.HadContextType);
E.setHadSymbolType(Relevance.HadSymbolType);
E.setTypeMatchesPreferred(Relevance.TypeMatchesPreferred);
E.setFilterLength(Relevance.FilterLength);
return Evaluate(E);
}
// Produces an integer that sorts in the same order as F.
// That is: a < b <==> encodeFloat(a) < encodeFloat(b).
static uint32_t encodeFloat(float F) {

View File

@ -77,6 +77,7 @@ struct SymbolQualitySignals {
void merge(const CodeCompletionResult &SemaCCResult);
void merge(const Symbol &IndexResult);
// FIXME(usx): Rename to evaluateHeuristics().
// Condense these signals down to a single number, higher is better.
float evaluate() const;
};
@ -136,6 +137,10 @@ struct SymbolRelevanceSignals {
// Whether the item matches the type expected in the completion context.
bool TypeMatchesPreferred = false;
/// Length of the unqualified partial name of Symbol typed in
/// CompletionPrefix.
unsigned FilterLength = 0;
/// Set of derived signals computed by calculateDerivedSignals(). Must not be
/// set explicitly.
struct DerivedSignals {
@ -161,6 +166,8 @@ llvm::raw_ostream &operator<<(llvm::raw_ostream &,
/// Combine symbol quality and relevance into a single score.
float evaluateSymbolAndRelevance(float SymbolQuality, float SymbolRelevance);
float evaluateDecisionForest(const SymbolQualitySignals &Quality,
const SymbolRelevanceSignals &Relevance);
/// TopN<T> is a lossy container that preserves only the "best" N elements.
template <typename T, typename Compare = std::greater<T>> class TopN {
public:

View File

@ -167,6 +167,26 @@ opt<CodeCompleteOptions::CodeCompletionParse> CodeCompletionParse{
Hidden,
};
opt<CodeCompleteOptions::CodeCompletionRankingModel> RankingModel{
"ranking-model",
cat(Features),
desc("Model to use to rank code-completion items"),
values(clEnumValN(CodeCompleteOptions::Heuristics, "heuristics",
"Use hueristics to rank code completion items"),
clEnumValN(CodeCompleteOptions::DecisionForest, "decision_forest",
"Use Decision Forest model to rank completion items")),
init(CodeCompleteOptions().RankingModel),
Hidden,
};
opt<bool> DecisionForestBase{
"decision-forest-base",
cat(Features),
desc("Base for exponentiating the prediction from DecisionForest."),
init(CodeCompleteOptions().DecisionForestBase),
Hidden,
};
// FIXME: also support "plain" style where signatures are always omitted.
enum CompletionStyleFlag { Detailed, Bundled };
opt<CompletionStyleFlag> CompletionStyle{
@ -739,6 +759,8 @@ clangd accepts flags on the commandline, and in the CLANGD_FLAGS environment var
CCOpts.EnableFunctionArgSnippets = EnableFunctionArgSnippets;
CCOpts.AllScopes = AllScopesCompletion;
CCOpts.RunParser = CodeCompletionParse;
CCOpts.RankingModel = RankingModel;
CCOpts.DecisionForestBase = DecisionForestBase;
RealThreadsafeFS TFS;
std::vector<std::unique_ptr<config::Provider>> ProviderStack;

View File

@ -10,7 +10,6 @@
#include "ClangdServer.h"
#include "CodeComplete.h"
#include "Compiler.h"
#include "CompletionModel.h"
#include "Matchers.h"
#include "Protocol.h"
#include "Quality.h"
@ -163,14 +162,38 @@ Symbol withReferences(int N, Symbol S) {
return S;
}
TEST(DecisionForestRuntime, SanityTest) {
using Example = clangd::Example;
using clangd::Evaluate;
Example E1;
E1.setContextKind(ContextKind::CCC_ArrowMemberAccess);
Example E2;
E2.setContextKind(ContextKind::CCC_SymbolOrNewName);
EXPECT_GT(Evaluate(E1), Evaluate(E2));
TEST(DecisionForestRankingModel, NameMatchSanityTest) {
clangd::CodeCompleteOptions Opts;
Opts.RankingModel = CodeCompleteOptions::DecisionForest;
auto Results = completions(
R"cpp(
struct MemberAccess {
int ABG();
int AlphaBetaGamma();
};
int func() { MemberAccess().ABG^ }
)cpp",
/*IndexSymbols=*/{}, Opts);
EXPECT_THAT(Results.Completions,
ElementsAre(Named("ABG"), Named("AlphaBetaGamma")));
}
TEST(DecisionForestRankingModel, ReferencesAffectRanking) {
clangd::CodeCompleteOptions Opts;
Opts.RankingModel = CodeCompleteOptions::DecisionForest;
constexpr int NumReferences = 100000;
EXPECT_THAT(
completions("int main() { clang^ }",
{ns("clangA"), withReferences(NumReferences, func("clangD"))},
Opts)
.Completions,
ElementsAre(Named("clangD"), Named("clangA")));
EXPECT_THAT(
completions("int main() { clang^ }",
{withReferences(NumReferences, ns("clangA")), func("clangD")},
Opts)
.Completions,
ElementsAre(Named("clangA"), Named("clangD")));
}
TEST(CompletionTest, Limit) {