Merge remote-tracking branch 'origin/metal-coreml' into ios-dev

cpp/configs/gtp_human9d_search_example.cfg

@@ -1,6 +1,6 @@
 
 # This is an example config for configuring KataGo to attempt to imitate a 9d player and then
-# further improving the strength using KataGo's evaluations and search.
+# further improving the strength (to possibly modestly superhuman) using KataGo's evaluations and search.
 
 # By contrast, gtp_human5k_example.cfg uses only the raw neural net. Raw neural nets like the
 # HumanSL net without search do NOT play accurately enough to reach top human levels.
@@ -44,6 +44,8 @@ resignMinMovesPerBoardArea = 0.40
 # This is due to the strong bias of the human SL network.
 # You can reduce this number if you are on weaker hardware. It may reduce strength a bit but will still
 # provide a huge strength boost over using the humanSL network alone as in gtp_human5k_example.cfg
+# It's NOT recommended to reduce this below about 30-40 visits, since that will result in too few explorations of a variety of moves.
+# If you want to adjust strength, see humanSLChosenMovePiklLambda instead.
 maxVisits = 400   # 40 in gtp_human5k_example.cfg.
 
 # Having more than one thread speeds up search when visits are larger.
@@ -61,14 +63,17 @@ humanSLProfile = preaz_9d    # was preaz_5k in gtp_human5k_example.cfg
 humanSLChosenMoveProp = 1.0
 humanSLChosenMoveIgnorePass = true
 
-# When a move starts to lose more than 0.06 utility (several percent winrate), downweight it.
-# Increase this number to reduce the strength gain and use the human SL policy move and KataGo's evaluations less.
+# When a move starts to lose more than 0.08 utility (several percent winrate), downweight it.
+# Increase this number to reduce the strength and use the human SL policy more and KataGo's evaluations less.
+# Decrease this number a little more to improve strength even further and play less human-like.
+#   (although below 0.02 you probably are better off going back to a normal KataGo config and scaling visits).
 # Since this uses KataGo's judgment, even at large values and with a weak-ranked humanSLProfile
 # this may still produce a very strong player.
-# To calibrate for some but less strength gain, it will take experimentation - in addition to increasing this value a lot,
+# To calibrate for some but less strength gain, it will take experimentation.
+# In addition to increasing this value a lot, e.g. (0.2, 0.5, 1.0, 2.0,...)
 # you can also try using old/small KataGo nets (e.g. b6c96, b10c128), reducing visits (though reducing below about 30-40
-# is not recommended), or using the humanSLModel itself as the main "-model".
-humanSLChosenMovePiklLambda = 0.06  # was 100000000 in gtp_human5k_example.cfg.
+# is NOT recommended), or using the humanSLModel itself as the main "-model".
+humanSLChosenMovePiklLambda = 0.08  # was 100000000 in gtp_human5k_example.cfg.
 
 # Spend 80% of visits to explore humanSL moves to ensure they get evaluations to use with humanSLChosenMovePiklLambda
 humanSLRootExploreProbWeightless = 0.8  # was 0 in gtp_human5k_example.cfg.

cpp/main.cpp

@@ -210,11 +210,11 @@ int main(int argc, const char* const* argv) {
 
 
 string Version::getKataGoVersion() {
-  return string("1.15.2-coreml1");
+  return string("1.15.3");
 }
 
 string Version::getKataGoVersionForHelp() {
-  return string("KataGo v1.15.2-coreml1");
+  return string("KataGo v1.15.3");
 }
 
 string Version::getKataGoVersionFullInfo() {

cpp/neuralnet/coremlbackend.swift

@@ -32,24 +32,26 @@ public class CoreMLBackend {
         return "KataGoModel\(xLen)x\(yLen)fp\(precision)\(encoder)"
     }
 
-    let model: KataGoModel
+    var model: KataGoModel
     let xLen: Int
     let yLen: Int
     public let version: Int32
     let numSpatialFeatures: Int
     let numGlobalFeatures: Int
     let numMetaFeatures: Int
     let metaEncoderVersion: Int
+    let modelName: String
 
     var spatialSize: Int {
         numSpatialFeatures * yLen * xLen
     }
 
-    init(model: MLModel, xLen: Int, yLen: Int, metaEncoderVersion: Int) {
+    init(model: MLModel, xLen: Int, yLen: Int, metaEncoderVersion: Int, modelName: String) {
         self.model = KataGoModel(model: model)
         self.xLen = xLen
         self.yLen = yLen
         self.metaEncoderVersion = metaEncoderVersion
+        self.modelName = modelName
 
         // The model version must be at least 8.
         self.version = model.version
@@ -115,8 +117,8 @@ public class CoreMLBackend {
 
             let inputBatch = KataGoModelInputBatch(inputArray: inputArray)
             let options = MLPredictionOptions()
-            let outputBatch = try! model.prediction(from: inputBatch, options: options)
 
+            let outputBatch = safelyPredict(from: inputBatch, options: options)!
             assert(outputBatch.count == batchSize)
 
             outputBatch.outputArray.enumerated().forEach { index, output in
@@ -148,6 +150,37 @@ public class CoreMLBackend {
             }
         }
     }
+
+    func safelyPredict(from inputBatch: KataGoModelInputBatch,
+                       options: MLPredictionOptions) -> KataGoModelOutputBatch? {
+        if let prediction = try? model.prediction(from: inputBatch, options: options) {
+            return prediction
+        }
+
+        let computeUnits = model.model.configuration.computeUnits
+
+        for mustCompile in [false, true] {
+            if let prediction = compileAndPredict(with: computeUnits, from: inputBatch, options: options, mustCompile: mustCompile) {
+                return prediction
+            }
+        }
+
+        return nil
+    }
+
+    private func compileAndPredict(with computeUnits: MLComputeUnits,
+                                   from inputBatch: KataGoModelInputBatch,
+                                   options: MLPredictionOptions,
+                                   mustCompile: Bool) -> KataGoModelOutputBatch? {
+        if let mlmodel = KataGoModel.compileBundleMLModel(modelName: modelName, computeUnits: computeUnits, mustCompile: mustCompile) {
+            model = KataGoModel(model: mlmodel)
+            if let outputBatch = try? model.prediction(from: inputBatch, options: options) {
+                return outputBatch
+            }
+        }
+
+        return nil
+    }
 }
 
 public func maybeCreateCoreMLBackend(condition: Bool = true,
@@ -162,15 +195,18 @@ public func maybeCreateCoreMLBackend(condition: Bool = true,
     // Get the model name.
     let modelName = CoreMLBackend.getModelName(xLen: xLen, yLen: yLen, useFP16: useFP16, metaEncoderVersion: metaEncoderVersion)
 
+    // Specify compute units.
+    let computeUnits: MLComputeUnits = useCpuAndNeuralEngine ? .cpuAndNeuralEngine : .all
+
     // Compile the model in Bundle.
-    let mlmodel = KataGoModel.compileBundleMLModel(modelName: modelName, useCpuAndNeuralEngine: useCpuAndNeuralEngine)
+    let mlmodel = KataGoModel.compileBundleMLModel(modelName: modelName, computeUnits: computeUnits)
 
     if let mlmodel {
         printError("CoreML backend \(serverThreadIdx): \(xLen)x\(yLen) useFP16 \(useFP16) metaEncoderVersion \(metaEncoderVersion) useCpuAndNeuralEngine \(useCpuAndNeuralEngine)");
         printError("CoreML backend \(serverThreadIdx): \(mlmodel.metaDescription)");
 
         // The CoreMLBackend object is created.
-        return CoreMLBackend(model: mlmodel, xLen: xLen, yLen: yLen, metaEncoderVersion: metaEncoderVersion)
+        return CoreMLBackend(model: mlmodel, xLen: xLen, yLen: yLen, metaEncoderVersion: metaEncoderVersion, modelName: modelName)
     } else {
         printError("Unable to compile bundle MLModel from model: \(modelName)")
         return nil

cpp/neuralnet/coremlmodel.swift

@@ -104,7 +104,7 @@ class KataGoModel {
         return bundleModelURL
     }
 
-    class func compileBundleMLModel(modelName: String, useCpuAndNeuralEngine: Bool) -> MLModel? {
+    class func compileBundleMLModel(modelName: String, computeUnits: MLComputeUnits, mustCompile: Bool = false) -> MLModel? {
         var mlmodel: MLModel?
 
         do {
@@ -114,7 +114,8 @@ class KataGoModel {
             // Compile MLModel
             mlmodel = try compileMLModel(modelName: modelName,
                                          modelURL: bundleModelURL,
-                                         useCpuAndNeuralEngine: useCpuAndNeuralEngine)
+                                         computeUnits: computeUnits,
+                                         mustCompile: mustCompile)
         } catch {
             printError("An error occurred: \(error)")
         }
@@ -225,9 +226,9 @@ class KataGoModel {
         try digest.write(to: savedDigestURL, atomically: true, encoding: .utf8)
     }
 
-    private class func loadModel(permanentURL: URL, modelName: String, useCpuAndNeuralEngine: Bool) throws -> MLModel {
+    private class func loadModel(permanentURL: URL, modelName: String, computeUnits: MLComputeUnits) throws -> MLModel {
         let configuration = MLModelConfiguration()
-        configuration.computeUnits = useCpuAndNeuralEngine ? .cpuAndNeuralEngine : .all
+        configuration.computeUnits = computeUnits
         configuration.modelDisplayName = modelName
         printError("Creating CoreML model with contents \(permanentURL)")
         return try MLModel(contentsOf: permanentURL, configuration: configuration)
@@ -247,14 +248,14 @@ class KataGoModel {
         return savedDigestURL
     }
 
-    class func compileMLModel(modelName: String, modelURL: URL, useCpuAndNeuralEngine: Bool) throws -> MLModel {
+    class func compileMLModel(modelName: String, modelURL: URL, computeUnits: MLComputeUnits, mustCompile: Bool) throws -> MLModel {
         let permanentURL = try getMLModelCPermanentURL(modelName: modelName)
         let savedDigestURL = try getSavedDigestURL(modelName: modelName)
         let digest = try getDigest(modelURL: modelURL)
 
-        let shouldCompileModel = checkShouldCompileModel(permanentURL: permanentURL,
-                                                         savedDigestURL: savedDigestURL,
-                                                         digest: digest)
+        let shouldCompileModel = mustCompile || checkShouldCompileModel(permanentURL: permanentURL,
+                                                                        savedDigestURL: savedDigestURL,
+                                                                        digest: digest)
 
         if shouldCompileModel {
             try compileAndSaveModel(permanentURL: permanentURL,
@@ -265,7 +266,7 @@ class KataGoModel {
 
         return try loadModel(permanentURL: permanentURL,
                              modelName: modelName,
-                             useCpuAndNeuralEngine: useCpuAndNeuralEngine);
+                             computeUnits: computeUnits);
     }
 
     init(model: MLModel) {

cpp/search/searchresults.cpp

@@ -295,8 +295,21 @@ bool Search::getPlaySelectionValues(
       maxValue = playSelectionValues[i];
   }
 
-  if(maxValue <= 1e-50)
-    return false;
+  if(maxValue <= 1e-50) {
+    //If we reached this point we have nonzero many children but the children are all weightless.
+    //In that case, at least set each one to be weighted by its policy.
+    for(int i = 0; i<numChildren; i++) {
+      playSelectionValues[i] = std::max(0.0,(double)policyProbs[getPos(locs[i])]);
+    }
+    //Recompute max
+    for(int i = 0; i<numChildren; i++) {
+      if(playSelectionValues[i] > maxValue)
+        maxValue = playSelectionValues[i];
+    }
+    if(maxValue <= 1e-50) {
+      return false;
+    }
+  }
 
   //Sanity check - if somehow we had more than this, something must have overflowed or gone wrong
   assert(maxValue < 1e40);

cpp/tests/results/analysis/badoverride.txt.log

@@ -22,7 +22,7 @@ rootSymmetryPruning = false
 trtUseFP16 = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/analysis/badoverride.txt.stderr

@@ -22,7 +22,7 @@ rootSymmetryPruning = false
 trtUseFP16 = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/analysis/basic.txt.log

@@ -22,7 +22,7 @@ rootSymmetryPruning = false
 trtUseFP16 = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/analysis/basic.txt.stderr

@@ -22,7 +22,7 @@ rootSymmetryPruning = false
 trtUseFP16 = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/analysis/basic_sidetomove.stderr

@@ -22,7 +22,7 @@ rootSymmetryPruning = false
 trtUseFP16 = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/analysis/basic_sidetomove.txt.log

@@ -22,7 +22,7 @@ rootSymmetryPruning = false
 trtUseFP16 = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/analysis/humansl_sidetomove.stderr

@@ -23,7 +23,7 @@ rootSymmetryPruning = false
 trtUseFP16 = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/analysis/pvvisits.txt.log

@@ -22,7 +22,7 @@ rootSymmetryPruning = false
 trtUseFP16 = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/analysis/pvvisits.txt.stderr

@@ -22,7 +22,7 @@ rootSymmetryPruning = false
 trtUseFP16 = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/analysis/pvvisits_nograph.stderr

@@ -22,7 +22,7 @@ trtUseFP16 = false
 useGraphSearch = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/analysis/pvvisits_nograph.txt.log

@@ -22,7 +22,7 @@ trtUseFP16 = false
 useGraphSearch = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/analysis/symmetry.txt.log

@@ -22,7 +22,7 @@ rootSymmetryPruning = false
 trtUseFP16 = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/analysis/symmetry.txt.stderr

@@ -22,7 +22,7 @@ rootSymmetryPruning = false
 trtUseFP16 = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/analysis/symmetry_with_pruning.stderr

@@ -22,7 +22,7 @@ rootSymmetryPruning = true
 trtUseFP16 = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/analysis/symmetry_with_pruning.txt.log

@@ -22,7 +22,7 @@ rootSymmetryPruning = true
 trtUseFP16 = false
 
 : Analysis Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : nnRandSeed0 = analysisTest
 : After dedups: nnModelFile0 = tests/models/g170-b6c96-s175395328-d26788732.bin.gz useFP16 false useNHWC false
 : Initializing neural net buffer to be size 19 * 19 allowing smaller boards

cpp/tests/results/gtp/avoidcorners.log

@@ -37,7 +37,7 @@ searchRandSeed = forTesting
 trtUseFP16 = false
 
 : GTP Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : Using TrompTaylor rules initially, unless GTP/GUI overrides this
 : Using 1 CPU thread(s) for search
 : Added 48 shapes to penalize repeats for bot 0 from tests/data/cornermoves.sgf

cpp/tests/results/gtp/avoidcorners.stderr

@@ -1,4 +1,4 @@
-KataGo v1.15.2
+KataGo v1.15.3
 Using TrompTaylor rules initially, unless GTP/GUI overrides this
 Initializing board with boardXSize 19 boardYSize 19
 Loaded config configs/gtp_example.cfg and/or command-line and query overrides

cpp/tests/results/gtp/basic.txt.log

@@ -30,7 +30,7 @@ searchRandSeed = forTesting
 trtUseFP16 = false
 
 : GTP Engine starting...
-: KataGo v1.15.2
+: KataGo v1.15.3
 : Using TrompTaylor rules initially, unless GTP/GUI overrides this
 : Using 1 CPU thread(s) for search
 : nnRandSeed0 = forTesting

cpp/tests/results/gtp/basic.txt.stderr

@@ -1,4 +1,4 @@
-KataGo v1.15.2
+KataGo v1.15.3
 Using TrompTaylor rules initially, unless GTP/GUI overrides this
 Initializing board with boardXSize 19 boardYSize 19
 Loaded config configs/gtp_example.cfg and/or command-line and query overrides