On the hunt for a better representation

3 files changed, 55 insertions, 50 deletions

diff --git a/extract.sh b/extract.sh
index 4ded1c0..60e01f9 100755
--- a/extract.sh
+++ b/extract.sh
@@ -10,38 +10,40 @@ and (result != '1/2-1/2');"
 
 extract() {
 		cd data
+		if [ ! -f "$db_file" ]; then
+				wget "https://www.playtak.com/games_anon.db"
+		fi
 		for size in 5 6; do
 				echo Extracing games of size "$size"...
 				sqlite3 "$db_file" "$(query $1 $size)" | shuf > "playtak-$size"
 		done
 }
 
-if [ ! -f "$db_file" ]; then
-		wget "https://www.playtak.com/games_anon.db"
-fi
-
-extract notation,result
+process() {
+		echo -e "Beginning to process data\n"
+		for i in 5 6; do
+				echo "Size $i..."
+				./pptdb "$i" "data/playtak-$i" > "data/check-$i"
+				tail -n21 "data/check-$i"
+				echo Stripping overflows and illegal games...
+				grep -Fvxf "data/check-$i" "data/playtak-$i" > "data/good-playtak-$i"
+				echo -n "Generating training data... "
+				./pptdb "$i" "data/good-playtak-$i" generate
+				echo "Shuffling data..."
+				shuf "data/training-$i.csv" > "data/shuf-$i.csv"
+				mv "data/shuf-$i.csv" "data/training-$i.csv"
+				echo
+		done
+}
 
-echo Preparing pptdb
+# extract notation,result
 if [ ! -f "pptdb" ]; then
-	 make pptdb
+		echo Preparing pptdb
+		make pptdb
 fi
-echo -e "Beginning to process data\n"
-
-for i in 5 6; do
-		echo "Size $i..."
-		./pptdb "$i" "data/playtak-$i" > "data/check-$i"
-		tail -n21 "data/check-$i"
-		echo Stripping overflows and illegal games...
-		grep -Fvxf "data/check-$i" "data/playtak-$i" > "data/good-playtak-$i"
-		echo -n "Generating training data... "
-		./pptdb "$i" "data/good-playtak-$i" generate
-		echo "Shuffling data..."
-		shuf "data/training-$i.csv" > "data/shuf-$i.csv"
-		mv "data/shuf-$i.csv" "data/training-$i.csv"
-		echo
-done
 
+# process
 
-shuf -n50000 data/good-playtak-5 > data/smalltak-5 && ./pptdb 5 data/smalltak-5 generate && \
-head -n1 data/training-5.csv > t && tail -n+2 data/training-5.csv > test.csv && shuf test.csv >> t && mv t test.csv
+make pptdb
+shuf -n10000 data/good-playtak-5 > data/smalltak-5 && ./pptdb 5 data/smalltak-5 generate && \
+head -n1 data/training-5.csv > t && tail -n+2 data/training-5.csv > data/train5.csv && shuf data/train5.csv >> t && mv t data/train5.csv
diff --git a/src/pptdb.c b/src/pptdb.c
index 712aabc..7a26ffd 100644
--- a/src/pptdb.c
+++ b/src/pptdb.c
@@ -19,28 +19,29 @@ write_input(void) {
 	// Two layers of board_size * board_size:
 	float t; int h; uint16_t mask;
 	for (int k = 0; k < board_size * board_size; k++) {
-		// stacks encoded as balanced ternary, without caps and walls
+		// stacks encoded as balanced ternary
 		t = 0;
 		h = COUNT_AT(k);
+
 		if (h>0) {
 			mask = 1<<(h-1);
-			/* if (STONE_AT(k) != STONE_FLAT) h--; */
 			while (h-->0) {
-				t += (colours[k] & mask) ? +1.0 : -1.0;
-				t /= 3;
+				t += (colours[k] & mask) ? +1 : -1;
+				t/=3;
 				mask >>= 1;
 			}
 		}
 		fprintf(training_fh,"%.6f,",t);
 	}
-	int val;
+	float val;
 	for (int k = 0; k < board_size * board_size; k++) {
 		val = 0;
 		if (COUNT_AT(k)) {
-			if (STONE_AT(k) == STONE_STANDING) val = (colours[k] & 1) ? +1 : -1;
-			else if (STONE_AT(k) == STONE_CAPSTONE) val = (colours[k] & 1) ? +2 : -2;
+			if (STONE_AT(k) == STONE_STANDING) val = (colours[k] & 1) ? +1/3 : -1/3;
+			else if (STONE_AT(k) == STONE_CAPSTONE) val = (colours[k] & 1) ? +1.0 : -1.0;
+			else val = (colours[k] & 1) ? +2/3 : -2/3;
 		}
-		fprintf(training_fh,"%d,",val);
+		fprintf(training_fh,"%.6f,",val);
 	}
 }
 
diff --git a/src/train5.py b/src/train5.py
index 225068b..3718fd0 100644
--- a/src/train5.py
+++ b/src/train5.py
@@ -12,40 +12,42 @@ from tensorflow.keras.layers import Dropout
 from tensorflow.keras.layers import Dense
 
 model = Sequential([
-    Conv2D(5, kernel_size=3, padding='same', input_shape=(5, 5, 2)),
-    MaxPooling2D(pool_size=(2, 2), strides=None),
-    Activation("relu"),
-    Flatten(),
-    Dense(10, activation="relu"),
-    Dense(8, activation="relu"),
-    Dense(1, activation="sigmoid")
-
-    # Dense(40, activation="relu", input_shape=(52,)),
+    # Conv2D(5, kernel_size=5, padding='same', input_shape=(5, 5, 2)),
+    # MaxPooling2D(pool_size=2, strides=2),
+    # Activation("relu"),
+    # Flatten(),
+    # Dense(25, activation="relu"),
     # Dense(10, activation="relu"),
-    # Dense(5, activation="relu"),
     # Dense(1, activation="sigmoid")
+
+    Dense(30, activation="relu", input_shape=(52,)),  # 30
+    Dense(10, activation="relu"),                     # 10
+    Dense(8, activation="relu"),                      # 8
+    Dense(1, activation="sigmoid")
 ])
 
 model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy'])
 
 model.summary()
 
-csv = pd.read_csv("data/test.csv")
+csv = pd.read_csv("data/train5.csv")
 
-training_data = csv.copy().head(5000)
+training_data = csv.copy().head(40000)
 training_outcome = training_data.pop('Outcome')
 
 training_input = training_data.copy()
-training_input = training_input.drop(columns=training_data.keys()[0:2])
+# training_input = training_input.drop(columns=training_data.keys()[0:2])
 train = np.array(training_input)
-train = train.reshape((training_input.shape[0], 5, 5, 2))
+# train = train.reshape((training_input.shape[0], 5, 5, 2))
 
-model.fit(train, training_outcome, epochs=20)
+model.fit(train, training_outcome, epochs=100)
 
-test_data = csv.copy().head(10000)
-test_data = test_data.drop(columns=test_data.keys()[0:2])
+test_data = csv.copy().tail(50000)
+# test_data = test_data.drop(columns=test_data.keys()[0:2])
 test_outcome = test_data.pop('Outcome')
 test_input = np.array(test_data)
-test_input.reshape((test_data.shape[0], 5, 5, 2))
+# test_input = test_input.reshape((test_data.shape[0], 5, 5, 2))
 test_loss, test_acc = model.evaluate(test_input,  test_outcome, verbose=2)
 print('\nTest accuracy:', test_acc)
+
+# Test accuracy: 0.67448