/
MrZeroTreeSimple.py
executable file
·336 lines (302 loc) · 13.5 KB
/
MrZeroTreeSimple.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
#!/usr/bin/env python3
# -*- coding: UTF-8 -*-
from Util import log,calc_score
from Util import ORDER_DICT,ORDER_DICT2,ORDER_DICT5,SCORE_DICT,INIT_CARDS
from MrRandom import MrRandom
from MrGreed import MrGreed
from ScenarioGenerator.ScenarioGen import ScenarioGen
from MCTS.mcts import mcts
from OfflineInterface import OfflineInterface
import torch
import torch.nn.functional as F
import copy,math
class GameState():
def __init__(self,cards_lists,score_lists,cards_on_table,play_for):
self.cards_lists=cards_lists
self.cards_on_table=cards_on_table
self.score_lists=score_lists
self.play_for=play_for
#decide cards_dicts, suit and pnext
self.cards_dicts=[MrGreed.gen_cards_dict(i) for i in self.cards_lists]
if len(self.cards_on_table)==1:
self.suit="A"
else:
self.suit=self.cards_on_table[1][0]
self.pnext=(self.cards_on_table[0]+len(self.cards_on_table)-1)%4
self.remain_card_num=sum([len(i) for i in self.cards_lists])
def getCurrentPlayer(self):
if (self.pnext-self.play_for)%2==0:
return 1
else:
return -1
def getPossibleActions(self):
return MrGreed.gen_legal_choice(self.suit,self.cards_dicts[self.pnext],self.cards_lists[self.pnext])
def takeAction(self,action):
#log(action)
neo_state=copy.deepcopy(self)
neo_state.cards_lists[neo_state.pnext].remove(action)
neo_state.cards_dicts[neo_state.pnext][action[0]].remove(action)
neo_state.remain_card_num-=1
neo_state.cards_on_table.append(action)
#log(neo_state.cards_on_table)
#input()
assert len(neo_state.cards_on_table)<=5
if len(neo_state.cards_on_table)<5:
neo_state.pnext=(neo_state.pnext+1)%4
if len(neo_state.cards_on_table)==2:
neo_state.suit=neo_state.cards_on_table[1][0]
else:
#decide pnext
score_temp=-1024
for i in range(4):
if neo_state.cards_on_table[i+1][0]==neo_state.cards_on_table[1][0] and ORDER_DICT2[neo_state.cards_on_table[i+1][1]]>score_temp:
winner=i #in relative order
score_temp=ORDER_DICT2[neo_state.cards_on_table[i+1][1]]
neo_state.pnext=(neo_state.cards_on_table[0]+winner)%4
#clear scores
neo_state.score_lists[neo_state.pnext]+=[c for c in neo_state.cards_on_table[1:] if c in SCORE_DICT]
#clean table
neo_state.cards_on_table=[neo_state.pnext,]
neo_state.suit='A'
return neo_state
def isTerminal(self):
if self.remain_card_num==0:
return True
else:
return False
def getReward(self):
#assert sum([len(i) for i in self.score_lists])==16
scores=[calc_score(self.score_lists[(self.play_for+i)%4]) for i in range(4)]
return scores[0]+scores[2]-scores[1]-scores[3]
print_level=0
BETA=0.2 #for pareparing train data
MCTS_EXPL=30
class MrZeroTreeSimple(MrRandom):
def __init__(self,room=0,place=0,name="default",pv_net=None,device=None,train_mode=False,
sample_b=10,sample_k=1,mcts_b=20,mcts_k=2):
MrRandom.__init__(self,room,place,name)
if isinstance(device,str):
self.device=torch.device(device)
else:
self.device=device
if isinstance(pv_net,str):
self.load_pv_net(net_para_loc=pv_net)
else:
self.pv_net=pv_net
self.sample_b=sample_b
self.sample_k=sample_k
self.mcts_b=mcts_b
self.mcts_k=mcts_k
self.train_mode=train_mode
if self.train_mode:
self.train_datas=[]
def load_pv_net(self,net_para_loc=None):
from MrZ_NETs import PV_NET_2
self.pv_net=PV_NET_2()
try:
self.pv_net.load_state_dict(torch.load(net_para_loc,map_location=self.device))
except FileNotFoundError:
self.pv_net.load_state_dict(torch.load("../"+net_para_loc,map_location=self.device))
self.pv_net.to(self.device)
def cards_lists_oh(cards_lists,place):
"""
return a 208-length one hot, in raletive order
the order is [me,me+1,me+2,me+3]
"""
oh=torch.zeros(52*4)#,dtype=torch.uint8)
for i in range(4):
for c in cards_lists[(place+i)%4]:
oh[52*i+ORDER_DICT[c]]=1
return oh
def score_lists_oh(score_lists,place):
"""
return a 64-length one hot, in relative order
the order is [me,me+1,me+2,me+3]
"""
oh=torch.zeros(16*4)#,dtype=torch.uint8)
for i in range(4):
for c in score_lists[(place+i)%4]:
oh[16*i+ORDER_DICT5[c]]=1
return oh
def four_cards_oh(cards_on_table,place):
"""
return a 156-legth oh, in anti-relative order
the order is [me-1,me-2,me-3]
"""
assert (cards_on_table[0]+len(cards_on_table)-1)%4==place
"""oh=torch.zeros(52*3)
for i,c in enumerate(cards_on_table[:0:-1]):
oh[52*i+ORDER_DICT[c]]=1"""
oh=torch.zeros(54*3)
for i,c in enumerate(cards_on_table[:0:-1]):
index=54*i+ORDER_DICT[c]#TODO +1 !!!
oh[index-1:index+2]=1
"""oh=torch.zeros(54*3+20*4)#,dtype=torch.uint8)
for i,c in enumerate(cards_on_table[:0:-1]):
index=54*i+ORDER_DICT[c]
oh[index-1:index+2]=1
oh[54*3+20*len(cards_on_table)-13:54*3+20*len(cards_on_table)]=1"""
return oh
def prepare_ohs(cards_lists,cards_on_table,score_lists,place):
oh_card=MrZeroTreeSimple.cards_lists_oh(cards_lists,place)
oh_score=MrZeroTreeSimple.score_lists_oh(score_lists,place)
oh_table=MrZeroTreeSimple.four_cards_oh(cards_on_table,place)
return torch.cat([oh_card,oh_score,oh_table])
def pv_policy(self,state):
if state.isTerminal():
return state.getReward()
else:
netin=MrZeroTreeSimple.prepare_ohs(state.cards_lists,state.cards_on_table,state.score_lists,state.pnext)
with torch.no_grad():
_,v=self.pv_net(netin.to(self.device))
return v.item()*state.getCurrentPlayer()+state.getReward()
def pick_a_card(self):
#input("in pick a card")
#确认桌上牌的数量和自己坐的位置相符
assert (self.cards_on_table[0]+len(self.cards_on_table)-1)%4==self.place
#utility datas
suit=self.decide_suit() #inherited from MrRandom
cards_dict=MrGreed.gen_cards_dict(self.cards_list)
#如果别无选择
if cards_dict.get(suit)!=None and len(cards_dict[suit])==1:
choice=cards_dict[suit][0]
if print_level>=1:
log("I have no choice but %s."%(choice))
return choice
if len(self.cards_list)==1:
if print_level>=1:
log("There is only one card left.")
return self.cards_list[0]
if print_level>=1:
log("my turn: %s, %s, %s"%(self.cards_on_table,self.cards_list,self.scores))
#生成Scenario
sce_num=self.sample_b+int(self.sample_k*len(self.cards_list))
sce_gen=ScenarioGen(self.place,self.history,self.cards_on_table,self.cards_list,number=sce_num)
cards_lists_list=[]
for cll in sce_gen:
cards_lists=[None,None,None,None]
cards_lists[self.place]=copy.copy(self.cards_list)
for i in range(3):
cards_lists[(self.place+i+1)%4]=cll[i]
cards_lists_list.append(cards_lists)
#MCTS并对Scenario平均
legal_choice=MrGreed.gen_legal_choice(suit,cards_dict,self.cards_list)
d_legal={c:0 for c in legal_choice}
searchnum=self.mcts_b+self.mcts_k*len(legal_choice)
for i,cards_lists in enumerate(cards_lists_list):
#initialize gamestate
gamestate=GameState(cards_lists,self.scores,self.cards_on_table,self.place)
#mcts
if self.mcts_k>=0:
searcher=mcts(iterationLimit=searchnum,rolloutPolicy=self.pv_policy,
explorationConstant=MCTS_EXPL)
searcher.search(initialState=gamestate)
for action,node in searcher.root.children.items():
d_legal[action]+=(node.totalReward/node.numVisits)/len(cards_lists_list)
elif self.mcts_k==-1:
input("not using this mode")
netin=MrZeroTreeSimple.prepare_ohs(cards_lists,self.cards_on_table,self.scores,self.place)
with torch.no_grad():
p,_=self.pv_net(netin.to(self.device))
p_legal=[(c,p[ORDER_DICT[c]]) for c in legal_choice]
p_legal.sort(key=lambda x:x[1],reverse=True)
d_legal[p_legal[0][0]]+=1
else:
raise Exception("reserved")
#挑选出最好的并返回
#d_legal={k:v/ for k,v in d_legal.items()}
best_choice=MrGreed.pick_best_from_dlegal(d_legal)
return best_choice
def pick_a_card_complete_info(self):
#确认桌上牌的数量和自己坐的位置相符
assert (self.cards_on_table[0]+len(self.cards_on_table)-1)%4==self.place
#initialize gamestate
#assert self.cards_list==self.cards_remain[self.place]
gamestate=GameState(self.cards_remain,self.scores,self.cards_on_table,self.place)
#mcts
suit=self.decide_suit()
cards_dict=MrGreed.gen_cards_dict(self.cards_list)
legal_choice=MrGreed.gen_legal_choice(suit,cards_dict,self.cards_list)
searchnum=self.mcts_b+self.mcts_k*len(legal_choice)
searcher=mcts(iterationLimit=searchnum,rolloutPolicy=self.pv_policy,
explorationConstant=MCTS_EXPL)
searcher.search(initialState=gamestate)
d_legal_temp={action: node.totalReward/node.numVisits for action,node in searcher.root.children.items()}
#save data for train
if self.train_mode:
value_max=max(d_legal_temp.values())
target_p=torch.zeros(52)
legal_mask=torch.zeros(52)
for k,v in d_legal_temp.items():
target_p[ORDER_DICT[k]]=math.exp(BETA*(v-value_max))
legal_mask[ORDER_DICT[k]]=1
target_p/=target_p.sum()
target_v=torch.tensor(value_max-gamestate.getReward())
netin=MrZeroTreeSimple.prepare_ohs(self.cards_remain,self.cards_on_table,self.scores,self.place)
self.train_datas.append((netin,target_p,target_v,legal_mask))
best_choice=MrGreed.pick_best_from_dlegal(d_legal_temp)
return best_choice
@staticmethod
def family_name():
return 'MrZeroTreeSimple'
BENCH_SMP_B=5
BENCH_SMP_K=0
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 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 prepare_data(pv_net,device_num,data_rounds,train_b,train_k):
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()
return zt[0].train_datas+zt[1].train_datas+zt[2].train_datas+zt[3].train_datas
if __name__=="__main__":
pass