one output (of course), Pade approximant of logistic for activation

1 files changed, 89 insertions, 0 deletions

diff --git a/include/ct1975.c b/include/ct1975.c
new file mode 100644
index 0000000..760fcd1
--- /dev/null
+++ b/include/ct1975.c
@@ -0,0 +1,89 @@
+#include "ct1975.h"
+
+float flattened[CONV_NUM+2];
+float dense1[DENSE1_NUM];
+float dense2[DENSE2_NUM];
+
+#define RELU(x) ((x) = ((x)<0)?0:(x))
+
+float
+evaluate_black_win(void) {
+	/* ------------------ *
+	 *  Convolution layer *
+	 * ------------------ */
+	// for each kernel
+	for (uint8_t kern = 0; kern < KERN_NUM; kern++) {
+		// the stride is 1, march across the board
+		for (uint8_t bx = 0; bx < KERN_OSIZE; bx++) {
+			for (uint8_t by = 0; by < KERN_OSIZE; by++) {
+				flattened[kern+KERN_NUM*(bx+by*KERN_OSIZE)] =
+					conv2d_biases[kern];
+				// Compute the convolution for this position
+				for (uint8_t ky = 0; ky < KERN_SIZE; ky++) {
+					for (uint8_t kx = 0; kx < KERN_SIZE; kx++) {
+						for (uint8_t c = 0; c < KERN_CHAN; c++) {
+							// Where we are on the board
+							const uint8_t loc = kx+by+(ky+bx)*board_size;
+							// Look up what's on the board at this location, and
+							// multiply it. For c=0 we have to do some extra work
+							float lookup = 0;
+							if (COUNT_AT(loc)>c) {
+								if (c==0) {
+									if (STONE_AT(loc) == STONE_STANDING) {
+										lookup = (colours[loc] & 1) ? +0.25 : -0.25;
+									} else if (STONE_AT(loc) == STONE_CAPSTONE) {
+										lookup = (colours[loc] & 1) ? +1.00 : -1.00;
+									} else {
+										lookup = (colours[loc] & 1) ? +0.50 : -0.50;
+									}
+								} else {
+									lookup = (colours[loc] & (1<<c)) ? +0.50 : -0.50;
+								}
+							}
+							flattened[kern+KERN_NUM*(bx+by*KERN_OSIZE)]
+								+= lookup*conv2d_weights[kern][ky][kx][c];
+						}
+					}
+				}
+				RELU(flattened[kern+KERN_NUM*(bx+by*KERN_OSIZE)]);
+			}
+		}
+	}
+	// Add input of flat counts
+	flattened[CONV_NUM]   = (float)(white_count & 127)/21.0;
+	flattened[CONV_NUM+1] = (float)(black_count & 127)/21.0;
+	/* ------------------ *
+	 *  First dense layer *
+	 * ------------------ */
+	for (uint8_t d1 = 0; d1 < DENSE1_NUM; d1++) {
+		dense1[d1] = dense1_biases[d1];
+		for (uint8_t fl = 0; fl < CONV_NUM+2; fl++) {
+			dense1[d1] += flattened[fl]*dense1_weights[d1][fl];
+		}
+		RELU(dense1[d1]);
+	}
+	/* ------------------- *
+	 *  Second dense layer *
+	 * ------------------- */
+	for (uint8_t d2 = 0; d2 < DENSE2_NUM; d2++) {
+		dense2[d2] = dense2_biases[d2];
+		for (uint8_t d1 = 0; d1 < DENSE1_NUM; d1++) {
+			dense2[d2] += dense1[d1]*dense2_weights[d2][d1];
+		}
+		RELU(dense2[d2]);
+	}
+	/* ------------- *
+	 *  Output layer *
+	 * ------------- */
+	float output = output_bias;
+	for (uint8_t d2 = 0; d2 < DENSE2_NUM; d2++) {
+		output += dense2[d2]*output_weights[d2];
+	}
+
+	// Truncated Pade approximant of logistic function
+	output = (12.0+output+50.0*output/(output*output+10.0))/24.0;
+	if (output > 1.0) return 1.0;
+	else if (output < 0.0) return 0.0;
+
+	return output;
+}