def try_mro():
from MrIf import MrIf
from OfflineInterface import OfflineInterface
import itertools,numpy
#netpara="MrO_Net_9_164609.pkl"
netpara="MrO_Net_15_395265.pkl"
o0=MrO(netpara,room=0,place=0,name='o0')
o2=MrO(netpara,room=0,place=2,name='o2')
f1=MrIf(room=0,place=1,name="if1")
f3=MrIf(room=0,place=3,name="if3")
offlineinterface=OfflineInterface([o0,f1,o2,f3],print_flag=False)
N1=64;N2=2
stats=[]
for k,l in itertools.product(range(N1),range(N2)):
if l==0:
cards=offlineinterface.shuffle()
else:
cards=cards[39:52]+cards[0:39]
offlineinterface.shuffle(cards=cards)
for i,j in itertools.product(range(13),range(4)):
offlineinterface.step()
stats.append(offlineinterface.clear())
if l==N2-1:
print("%4d"%(sum([j[0]+j[2]-j[1]-j[3] for j in stats[-N2:]])),end=" ",flush=True)
offlineinterface.prepare_new()
s_temp=[j[0]+j[2]-j[1]-j[3] for j in stats]
print("%.2f %.2f"%(numpy.mean(s_temp),numpy.sqrt(numpy.var(s_temp)/(len(s_temp)-1)),))
def gen_data_for_o(N1=2,N2=1,save=False):
from OfflineInterface import OfflineInterface
import pickle
global for_o
for_o=([],[],[],[])
g=[MrGreed(room=0,place=i,name='greed%d'%(i)) for i in range(4)]
offlineinterface=OfflineInterface(g,print_flag=False)
stats=[]
for k,l in itertools.product(range(N1),range(N2)):
if l==0:
cards=offlineinterface.shuffle()
else:
cards=cards[39:52]+cards[0:39]
offlineinterface.shuffle(cards=cards)
for i,j in itertools.product(range(13),range(4)):
offlineinterface.step()
stats.append(offlineinterface.clear())
print(".",end=" ",flush=True)
offlineinterface.prepare_new()
print("")
if save:
with open("Greed_%d.data"%(N1),'wb') as f:
pickle.dump(for_o,f)
log("saved")
return for_o
def optimize_target(paras):
"""
will be called by optimize_para to optimize parameters of MrGreed
should import:
from MrIf import MrIf
from OfflineInterface import OfflineInterface
import numpy
"""
print(paras,end=" ",flush=True)
g0=MrGreed(room=0,place=0,name='greed0')
g2=MrGreed(room=0,place=2,name='greed2')
f1=MrIf(room=0,place=1,name="if1")
f3=MrIf(room=0,place=3,name="if3")
for g in [g0,g2]:
g.SHORT_PREFERENCE=paras[0]*100
offlineinterface=OfflineInterface([g0,f1,g2,f3],print_flag=False)
N1=256;N2=2
stats=[]
for k,l in itertools.product(range(N1),range(N2)):
if l==0:
cards=offlineinterface.shuffle()
else:
cards=cards[39:52]+cards[0:39]
offlineinterface.shuffle(cards=cards)
for i,j in itertools.product(range(13),range(4)):
offlineinterface.step()
stats.append(offlineinterface.clear())
offlineinterface.prepare_new()
s_temp=[j[0]+j[2]-j[1]-j[3] for j in stats]
print("%.2f %.2f"%(numpy.mean(s_temp),numpy.sqrt(numpy.var(s_temp)/(len(s_temp)-1)),))
return numpy.mean(s_temp)
def benchmark_B(handsfile):
from MrIf import MrIf
from MrGreed import MrGreed
from MrZeroTree import MrZeroTree
from OfflineInterface import OfflineInterface, read_std_hands, play_a_test
ifs = [MrIf(room=255, place=i, name='I%d' % (i)) for i in range(4)]
gs = [MrGreed(room=255, place=i, name='G%d' % (i)) for i in range(4)]
zs = [
MrZeroTree(room=255,
place=i,
name='Z%d' % (i),
mcts_b=10,
mcts_k=2,
sample_b=-1,
sample_k=-2) for i in [0, 2]
]
I_GI = OfflineInterface([gs[0], ifs[1], gs[2], ifs[3]], print_flag=False)
I_ZG = OfflineInterface([zs[0], gs[1], zs[1], gs[3]], print_flag=False)
hands = read_std_hands(handsfile)
stats = []
for k, hand in hands:
stats.append(play_a_test(I_ZG, hand, 2))
print("%4d" % (stats[-1], ), end=" ", flush=True)
else:
print("")
log("benchmark result: %.2f %.2f" %
(numpy.mean(stats), numpy.sqrt(numpy.var(stats) / (len(stats) - 1))))
def benchmark():
from MrRandom import MrRandom, Human
from MrIf import MrIf
from OfflineInterface import OfflineInterface
g = [MrGreed(room=0, place=i, name='greed%d' % (i)) for i in range(4)]
f = [MrIf(room=0, place=i, name="if%d" % (i)) for i in range(4)]
r = [MrRandom(room=0, place=i, name="random%d" % (i)) for i in range(4)]
rt = [
MrRandTree(room=0, place=i, name='randtree%d' % (i)) for i in range(4)
]
offlineinterface = OfflineInterface([f[0], g[1], f[2], g[3]],
print_flag=False)
N1 = 1024
N2 = 2
stats = []
log("%s vs. %s for %dx%d" %
(offlineinterface.players[0].family_name(),
offlineinterface.players[1].family_name(), N1, N2))
tik = time.time()
for k, l in itertools.product(range(N1), range(N2)):
if l == 0:
cards = offlineinterface.shuffle()
else:
cards = cards[39:52] + cards[0:39]
offlineinterface.shuffle(cards=cards)
for i, j in itertools.product(range(13), range(4)):
offlineinterface.step()
"""if i==7 and j==2:
global print_level
print_level=1
offlineinterface.print_flag=True
log("start outputs")"""
stats.append(offlineinterface.clear())
offlineinterface.prepare_new()
if l == N2 - 1:
print("%4d" %
(sum([j[0] + j[2] - j[1] - j[3] for j in stats[-N2:]]) / N2),
end=" ",
flush=True)
#print("%s"%(stats[-1]),end=" ",flush=True)
tok = time.time()
log("time consume: %ds" % (tok - tik))
for i in range(4):
s_temp = [j[i] for j in stats]
log("%dth player: %.2f %.2f" % (
i,
numpy.mean(s_temp),
numpy.sqrt(numpy.var(s_temp) / (len(s_temp) - 1)),
),
l=2)
s_temp = [j[0] + j[2] - j[1] - j[3] for j in stats]
log("%.2f %.2f" %
(numpy.mean(s_temp), numpy.sqrt(numpy.var(s_temp) /
(len(s_temp) - 1))))
def benchmark_transitivity(print_process=False):
from MrRandom import MrRandom
from MrIf import MrIf
from MrGreed import MrGreed
from MrRandTree import MrRandTree
from MrZ_NETs import PV_NET_2
from MrZeroTreeSimple import MrZeroTreeSimple
from OfflineInterface import OfflineInterface
import itertools,torch,random,inspect
device_bench=torch.device("cuda:0")
save_name_0="Zero-29th-25-11416629-720.pt"
state_dict_0=torch.load(save_name_0,map_location=device_bench)
pv_net_0=PV_NET_2()
pv_net_0.load_state_dict(state_dict_0)
pv_net_0.to(device_bench)
team0=[MrZeroTreeSimple(room=255,place=i,name='zts%d'%(i),pv_net=pv_net_0,device=device_bench,mcts_b=10,mcts_k=2,sample_b=9,sample_k=0) for i in [0,2]]
#team0=[MrRandTree(room=255,place=i,name='randtree%d'%(i)) for i in [0,2]]
#team1=[MrRandTree(room=255,place=i,name='randtree%d'%(i)) for i in [1,3]]
#team1=[MrGreed(room=255,place=i,name='greed%d'%(i)) for i in [1,3]]
#team1=[MrRandom(room=255,place=i,name='random%d'%(i)) for i in [1,3]]
team1=[MrIf(room=255,place=i,name='if%d'%(i)) for i in [1,3]]
interface=OfflineInterface([team0[0],team1[0],team0[1],team1[1]],print_flag=False)
N1=256;N2=2;
log("(%s+%s) v.s. (%s+%s) for %dx%d"%(interface.players[0].family_name(),interface.players[2].family_name(),
interface.players[1].family_name(),interface.players[3].family_name(),N1,N2))
if interface.players[0].family_name().startswith("MrZeroTree"):
log("mcts_b/k: %d/%d, sample_b/k: %d/%d"%(interface.players[0].mcts_b,interface.players[0].mcts_k,
interface.players[0].sample_b,interface.players[0].sample_k))
stats=[]
for k,l in itertools.product(range(N1),range(N2)):
if l==0:
cards=interface.shuffle()
else:
cards=cards[39:52]+cards[0:39]
interface.shuffle(cards=cards)
for i,j in itertools.product(range(13),range(4)):
interface.step()
#input("continue...")
stats.append(interface.clear())
interface.prepare_new()
if l==N2-1:
if print_process:
log("%2d %4d: %s"%(k,sum([j[0]+j[2]-j[1]-j[3] for j in stats[-N2:]])/N2,stats[-N2:]))
else:
print("%4d"%(sum([j[0]+j[2]-j[1]-j[3] for j in stats[-N2:]])/N2),end=" ",flush=True)
if (k+1)%(N1//4)==0 and l==N2-1:
bench_stat(stats,N2,None)
bench_stat(stats,N2,None)
def prepare_data_queue(pv_net, device_num, data_rounds, train_b, train_k,
data_queue):
input("not using")
device_train = torch.device("cuda:%d" % (device_num))
pv_net.to(device_train)
zt = [
MrZeroTreeSimple(room=0,
place=i,
name='zerotree%d' % (i),
pv_net=pv_net,
device=device_train,
train_mode=True,
mcts_b=train_b,
mcts_k=train_k) for i in range(4)
]
interface = OfflineInterface(zt, print_flag=False)
stats = []
for k in range(data_rounds):
cards = interface.shuffle()
for i in range(52):
interface.step_complete_info()
stats.append(interface.clear())
interface.prepare_new()
for i in range(4):
data_queue.put(zt[i].train_datas, block=False)
def benchmark(save_name, epoch, device_num, print_process=False):
"""
benchmark raw network against MrGreed
will be called by trainer
"""
import itertools, numpy
N1 = 512
N2 = 2
log("start benchmark against MrGreed for %dx%d" % (N1, N2))
zt = [
MrZeroTreeSimple(room=255,
place=i,
name='zerotree%d' % (i),
pv_net=save_name,
device="cuda:%d" % (device_num),
mcts_b=0,
mcts_k=1,
sample_b=BENCH_SMP_B,
sample_k=BENCH_SMP_K) for i in [0, 2]
]
g = [MrGreed(room=255, place=i, name='greed%d' % (i)) for i in [1, 3]]
interface = OfflineInterface([zt[0], g[0], zt[1], g[1]], print_flag=False)
stats = []
for k, l in itertools.product(range(N1), range(N2)):
if l == 0:
cards = interface.shuffle()
else:
cards = cards[39:52] + cards[0:39]
interface.shuffle(cards=cards)
for i, j in itertools.product(range(13), range(4)):
interface.step()
stats.append(interface.clear())
interface.prepare_new()
if print_process and l == N2 - 1:
print("%4d" %
(sum([j[0] + j[2] - j[1] - j[3] for j in stats[-N2:]]) / N2),
end=" ",
flush=True)
s_temp = [j[0] + j[2] - j[1] - j[3] for j in stats]
s_temp = [sum(s_temp[i:i + N2]) / N2 for i in range(0, len(s_temp), N2)]
log("benchmark at epoch %s's result: %.2f %.2f" %
(epoch, numpy.mean(s_temp),
numpy.sqrt(numpy.var(s_temp) / (len(s_temp) - 1))))
def benchmark(handsfile, print_process=False):
from MrGreed import MrGreed
from MrZeroTree import MrZeroTree
from OfflineInterface import OfflineInterface, read_std_hands, play_a_test
import torch, inspect
log_source(inspect.getsource(MrZeroTree.decide_rect_necessity))
#log_source(inspect.getsource(MrZeroTree.possi_rectify_pvnet))
zt0 = [
MrZeroTree(room=255,
place=i,
name='zerotree%d' % (i),
mcts_b=10,
mcts_k=2,
sample_b=-1,
sample_k=-2) for i in [0, 2]
]
team1 = [MrGreed(room=255, place=i, name='greed%d' % (i)) for i in [1, 3]]
interface = OfflineInterface([zt0[0], team1[0], zt0[1], team1[1]],
print_flag=False)
if interface.players[0].family_name().startswith("MrZeroTree"):
p0 = interface.players[0]
log("mcts_b/k: %d/%d, sample_b/k: %d/%d" %
(p0.mcts_b, p0.mcts_k, p0.sample_b, p0.sample_k))
hands = read_std_hands(handsfile)
N1 = len(hands)
N2 = 2
log("%s for %dx%d on %s" % (interface, N1, N2, zt0[0].device))
stats = []
for k, hand in hands:
stats.append(play_a_test(interface, hand, N2))
print("%4d" % (stats[-1], ), end=" ", flush=True)
if (k + 1) % (N1 // 4) == 0:
bench_stat(stats, N2)
bench_stat(stats, N2)
from MrIf import LOGFILE,log,cards_order,MrRandom,Human,MrIf
from OfflineInterface import OfflineInterface
if __name__=="__main__":
#创建一系列机器人对象
#机器人对象应当有如下方法
#receive_shuffle(self,cards) 接收洗牌
#pick_a_card(self,suit) 出哪张牌，但不更新数据结构，要等到offlineinterface调用机器人的pop_card才更新
#pop_card(self,which) 确认手牌打出后会被调用，更新手牌的数据结构
#但是offlineinterface还没有把历史信息传给机器人的功能，请自定义函数
random0=MrRandom(0,0,"random0")
random1=MrRandom(0,1,"random1")
random2=MrRandom(0,2,"random2")
random3=MrRandom(0,3,"random3")
if0=MrIf(0,0,"if0")
if1=MrIf(0,1,"if1")
if2=MrIf(0,2,"if2")
if3=MrIf(0,2,"if3")
#使用四个机器人初始化OfflineInterface（机器人也可以是Human，这样人就可以加入）
offlineinterface=OfflineInterface([if0,random1,if2,random3])
#发牌
offlineinterface.shuffle()
#或者指定发什么牌然后发牌
#offlineinterface.shuffle(cards=cards)
#打54张牌
for i in range(52):
offlineinterface.step()
#打印分数
log(offlineinterface.clear())
#准备重新开始
offlineinterface.prepare_new()