Merge pull request #66 from kobanium/develop

Develop

learn/convert/hcpe3_to_csa.py

@@ -0,0 +1,133 @@
+from cshogi import *
+from cshogi import CSA
+import numpy as np
+import sys
+import os
+import glob
+import math
+import argparse
+
+HuffmanCodedPosAndEval3 = np.dtype([
+    ('hcp', dtypeHcp), # 開始局面
+    ('moveNum', np.uint16), # 手数
+    ('result', np.uint8), # 結果（xxxxxx11:勝敗、xxxxx1xx:千日手、xxxx1xxx:入玉宣言、xxx1xxxx:最大手数）
+    ('opponent', np.uint8), # 対戦相手（0:自己対局、1:先手usi、2:後手usi）
+    ])
+MoveInfo = np.dtype([
+    ('selectedMove16', dtypeMove16), # 指し手
+    ('eval', dtypeEval), # 評価値
+    ('candidateNum', np.uint16), # 候補手の数
+    ])
+MoveVisits = np.dtype([
+    ('move16', dtypeMove16), # 候補手
+    ('visitNum', np.uint16), # 訪問回数
+    ])
+
+ENDGAME_SYMBOLS = {
+    1 : '%TORYO',
+    2 : '%TORYO',
+    4 : '%SENNICHITE',
+    9 : '%KACHI',
+    10: '%KACHI',
+    16: '%CHUDAN',
+}
+
+parser = argparse.ArgumentParser()
+parser.add_argument('hcpe3')
+parser.add_argument('csa')
+parser.add_argument('--range')
+parser.add_argument('--nyugyoku', action='store_true')
+parser.add_argument('--aoba', action='store_true')
+parser.add_argument('--out_v', action='store_true')
+parser.add_argument('--sort_visits', action='store_true')
+args = parser.parse_args()
+
+f = open(args.hcpe3, 'rb')
+
+if args.aoba:
+    sep = ','
+else:
+    sep = '\n'
+
+if args.range:
+    start_end = args.range.split(':')
+    if len(start_end) == 1:
+        start = int(start_end[0])
+        end = start + 1
+    else:
+        if start_end[0] == '':
+            start = 0
+        else:
+            start = int(start_end[0])
+        if start_end[1] == '':
+            end = sys.maxsize
+        else:
+            end = int(start_end[1])
+else:
+    start = 0
+    end = sys.maxsize
+
+board = Board()
+csa = CSA.Exporter(args.csa)
+v_pos_sum = 0
+move_sum = 0
+p = 0
+while p < end:
+    data = f.read(HuffmanCodedPosAndEval3.itemsize)
+    if len(data) == 0:
+        break
+    hcpe = np.frombuffer(data, HuffmanCodedPosAndEval3, 1)[0]
+    board.set_hcp(hcpe['hcp'])
+    assert board.is_ok()
+    move_num = hcpe['moveNum']
+    result = hcpe['result']
+    move_sum += move_num
+    need_output = p >= start and (not args.nyugyoku or result & 8 != 0)
+    if need_output:
+#        csa.info(board, comments=[f"moveNum={move_num},result={result},opponent={hcpe['opponent']}"])
+         print ("'move_num=" +str(move_num) +",result=" +str(result)+ ",opponent=" + str(hcpe['opponent']))
+    print (board)
+    for i in range(move_num):
+        move_info = np.frombuffer(f.read(MoveInfo.itemsize), MoveInfo, 1)[0]
+        candidate_num = move_info['candidateNum']
+        move_visits = np.frombuffer(f.read(MoveVisits.itemsize * candidate_num), MoveVisits, candidate_num)
+        move = board.move_from_move16(move_info['selectedMove16'])
+        if need_output:
+            if candidate_num > 0:
+                v_pos_sum += 1
+                if args.aoba:
+                    if args.out_v:
+                        v = 1.0 / (1.0 + math.exp(-move_info['eval'] * 0.0013226))
+                        comment = f"v={v:.3f},"
+                    else:
+                        comment = ''
+                    comment += f"{move_visits['visitNum'].sum()}"
+                    if args.sort_visits:
+                        move_visits = np.sort(move_visits, order='visitNum')[::-1]
+                    for move16, visit_num in zip(move_visits['move16'], move_visits['visitNum']):
+                        comment += ',' + move_to_csa(board.move_from_move16(move16)) + ',' + str(visit_num)
+                else:
+                    comment = '** ' + str(move_info['eval'] * (1 - board.turn * 2))
+            else:
+                if args.aoba or move_info['eval'] == 0:
+#                    comment = None
+                    comment = ""
+                else:
+                    comment = '** ' + str(move_info['eval'] * (1 - board.turn * 2))
+#            csa.move(move, comment=comment, sep=sep)
+#            csa.move(move, sep=sep)
+            if (board.turn) == 0:
+                sen = "+"
+            else:
+                sen = "-"
+            print (sen + move_to_csa(move) + ",'" + comment)
+        board.push(move)
+        assert board.is_ok()
+    if need_output:
+        print (ENDGAME_SYMBOLS[hcpe['result']])
+        print ("/")
+#        csa.endgame(ENDGAME_SYMBOLS[hcpe['result']])
+    p += 1
+print (p,move_sum,v_pos_sum)
+
+

learn/convert/hcpe_to_csa.py

@@ -0,0 +1,59 @@
+from cshogi import *
+import numpy as np
+import math
+
+import argparse
+
+parser = argparse.ArgumentParser()
+parser.add_argument('hcpe')
+#parser.add_argument('psv')
+args = parser.parse_args()
+
+hcpes = np.fromfile(args.hcpe, dtype=HuffmanCodedPosAndEval)
+#psvs = np.zeros(len(hcpes), PackedSfenValue)
+move_sum = 0
+res = [0,0,0]
+s_max = -999999
+s_min = +999999
+board = Board()
+for hcpe in hcpes:
+#for hcpe in zip(hcpes):
+    board.set_hcp(hcpe['hcp'])
+#    board.to_psfen(psv['sfen'])
+    score = hcpe['eval']
+    move16 = hcpe['bestMove16']
+    move_csa = move_to_csa(board.move_from_move16(move16))
+    #move_num = hcpe['moveNum']
+    result = hcpe['gameResult']
+    # gameResult -> 0: DRAW, 1: BLACK_WIN, 2: WHITE_WIN
+    #if board.turn == gameResult - 1:
+    #    psv['game_result'] = 1
+    #elif board.turn == 2 - gameResult:
+    #    psv['game_result'] = -1
+    move_sum += 1
+    res[result] += 1
+    if score > s_max:
+        s_max = score
+    if score < s_min:
+        s_min = score
+    print ("'move_sum=" +str(move_sum) +",result=" +str(result) +  ",move=" + move_csa + ",score=" + str(score))
+    print (board)
+
+    if (board.turn) == 0:
+        sen = "+"
+    else:
+        sen = "-"
+    v = 1.0 / (1.0 + math.exp(-score * 0.0013226))
+    comment = f"v={v:.3f},"
+    print (sen + move_csa + "," + comment)
+    s_res = "%SENNICHITE"
+    if result == 1:
+        s_res = "%-ILLEGAL_ACTION"
+    elif result == 2:
+        s_res = "%+ILLEGAL_ACTION"
+    print (s_res)
+    print ("/")
+
+#psvs.tofile(args.psv)
+print ("'move_sum=" +str(move_sum) +",result=" +str(result) +  ",move=" + move_csa + ",score=" + str(score))
+print ("'res[0]=" +str(res[0]) +",[1]=" +str(res[1]) + ",[2]=" + str(res[2]) + ",s_max="+str(s_max)+",s_min"+str(s_min))

learn/convert/readme.txt

@@ -0,0 +1,91 @@
+GCT�̊w�K�Ɏg�p�����f�[�^�Z�b�g�����J 
+https://tadaoyamaoka.hatenablog.com/entry/2021/05/06/223701
+
+python3 ./hcpe3_to_csa.py --aoba --out_v --sort_visits selfplay_gct-051.hcpe3 dummy.csa >> selfplay_gct-051.csa
+
+����AI�̎����m�[�g�F����̕��z���w�K ���̂Q
+https://tadaoyamaoka.hatenablog.com/entry/2021/03/16/235251
+���t�ǖʃt�H�[�}�b�g(hcpe)
+�ǖʒP�ʂɏo�͂��āA�s�v�ȋǖʁi���@�肪1�肵���Ȃ��ǖʁj�́A�o�͂��Ă��Ȃ������B
+�܂��A�΋ǂ̋��ڂ����m�ɂ킩��Ȃ�
+
+
+hcpe_to_csa.py �͈ȉ��̃f�B���N�g���ɒu���Ď��s
+
+$ python3 /home/yss/shogi/dlshogi_dr2_exhi/dlshogi/utils/hcpe_to_csa.py dlshogi_with_gct-024.hcpe > dlshogi_with_gct-024.csa
+
+��菟���A�̋ǖʂ�
+%-ILLEGAL_ACTION   (��肪��������w�����̂Ő�菟��)
+��菟���A�̋ǖʂ�
+%+ILLEGAL_ACTION   (��肪��������w�����̂Ō�菟��)
+�𗘗p�B
+
+
+
+
+
+
+
+selfplay_gct-001.hcpe3
+308133 ���� 3638986140 byte�ŁB
+308133 * 75 = 23,109,975
+2300���������x�B����75����x�ŁA15���ǖʁB
+
+
+
+
+dlshogi_with_gct-001.hcpe , 356250760 byte
+'move_sum=9375020,'res[0]=215305,[1]=4828653,[2]=4331062,s_max=30000,s_min-30000
+
+339MB�� 937���ǖʁA���łɃV���b�t������Ă���B�����ɖ߂��͍̂���B
+024 �܂�24�‚�
+9375020 * 24 = 225,000,480     2���ǖ�
+
+339 MB *24 = 8GB
+
+hcpe�ɂ͋ǖ�(���)�A���̎�A�]���l(-30000 <= v <= +30000),�΋nj���(0...���������A1...��菟���A2...��菟��)
+��4�‚̂݁B�萔�͂Ȃ��B
+����
+4828653��
+4331062�s
+ 215305��
+-------------
+(4828653 + 215305/2) / 9375020 = 0.5265  ��菟��0.526�ł���قǍ����Ȃ��B
+
+
+suisho3kai-001.csa  ,   55417414 byte / 1458353 �ǖ� = 38 byte/�ǖʁB�Ֆʂ�32byte���B
+842825��
+615528�s
+     0��
+-------------
+1458353 �ǖ�,  0.5779  ��菟����0.57�ō����B  �]���l(-31996 <= v <= +31997)�B�͈͂��Ⴄ�ȁB
+24�‚� 1458353 * 24 = 35,000,472 �ǖ�
+
+floodgate_2019-2021_r3500-001.hcpe , 11078976 byte
+155969 ��
+125196 �s
+ 10387 ��
+-----------
+291552 �ǖ�   0.5527  �]���l(-32767 <= v <= +32767) �B������������ɔ͈͂��Ⴄ
+24�‚� 291552 * 24 = 6,997,248 �ǖ�
+
+
+���v
+225,000,480   dlshogi_with_gct
+ 35,000,472   suisho3kai
+  6,997,248   floodgate_2019-2021_r3500
+--------------------
+266,998,200     2��6�疜�ǖʁB2021�N5��6���ł͖�6���B
+
+1�ǖ�100byte�Ƃ���25GB�B
+
+CSA�ɕϊ����ēǂݍ��񂾌�A
+266998238 �ǖʁB��⑽���B
+ 76.9 * 128GB = 98GB���g�p�B
+
+
+139910027 ��菟��
+121668247 ��菟��
+  5419964 ��������
+----------------------
+266998238  ��菟�� (0.534)

learn/yss.cpp

@@ -938,6 +938,15 @@ int shogi::LoadCSA()
 			ZERO_DB *pz = &zdb_one;
 			pz->handicap = get_handicap_from_board();
 			if ( pz->handicap && fGotekara==0 ) DEBUG_PRT("pz->handiacp=%d\n",pz->handicap);
+#if (GCT_SELF==1)
+			if ( pz->handicap != 0 ) DEBUG_PRT("");
+			for (int y=0;y<9;y++) for (int x=0;x<9;x++) {
+				pz->v_init_pos.push_back(init_ban[(y+1)*16+(x+1)]);
+			}
+			for (int i=0;i<7;i++) pz->v_init_pos.push_back(mo_m[i+1]);
+			for (int i=0;i<7;i++) pz->v_init_pos.push_back(mo_c[i+1]);
+			pz->v_init_pos.push_back(fGotekara);
+#endif
 		}
 
 		// csa形式のコメントを取り込む
@@ -1006,7 +1015,7 @@ int shogi::LoadCSA()
 							b0 = b1 = 0;
 						} else {
 							if ( getMoveFromCsaStr(&bz, &az, &tk, &nf, str)==0 ) DEBUG_PRT("");
-							int c = (tesuu+(pz->handicap!=0))&1;
+							int c = (tesuu + fGotekara)&1;
 							if ( is_pseudo_legalYSS((Move)pack_te(bz,az,tk,nf), (Color)(c==1) ) == false ) {
 								DEBUG_PRT("move Err %3d:%s\n",tesuu,str);
 							}
@@ -1167,7 +1176,7 @@ P-00AL
 			pz->moves = tesuu;
 			pz->result = ZD_DRAW;
 			pz->result_type = RT_NONE;
-			int is_gote_turn = (tesuu + (pz->handicap!=0))& 1;
+			int is_gote_turn = (tesuu + fGotekara) & 1;
 
 			if ( strstr(lpLine,"TORYO") ) {
 				if ( is_gote_turn ) {

learn/yss.h

@@ -385,7 +385,8 @@ class shogi {
 	void clear_init_ban();
 	void hirate_ban_init(int n);/*** 盤面を初期状態に戻す（駒落ち付き） ***/
 	int get_handicap_from_board();
-	void ban_saikousei(void);	// 現在の持ち駒、盤面の状態(ban_init)を元に盤面状態を再構成する。
+	void ban_saikousei();	// 現在の持ち駒、盤面の状態(ban_init)を元に盤面状態を再構成する。
+	void ban_saikousei_without_kiki();
 	int is_hirate_ban();	// 平手の盤面か判定する
 	void hanten_with_hash_kifu();
 
@@ -440,7 +441,8 @@ class shogi {
 	int wait_and_get_new_kif(int next_weight_n);
 	int add_a_little_from_archive();
 	int make_www_samples();
-
+	void get_piece_num_diff(bool bGoteTurn, int d[]);
+	void sum_pwv(double z, bool bGoteTurn, double sumd[]);
 
 
 	// fish関連