File size: 7,893 Bytes
fe56a4f cb658f7 fe56a4f cb658f7 fe56a4f cb658f7 fe56a4f cb658f7 fe56a4f cb658f7 fe56a4f cb658f7 fe56a4f cb658f7 fe56a4f |
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 |
from collections import deque
from Arena import Arena
from MCTS import MCTS
import numpy as np
from progress.bar import Bar
from quoridor.pytorch.NNet import AverageMeter
import time, os, sys
from pickle import Pickler, Unpickler
from random import shuffle
class Coach():
"""
This class executes the self-play + learning. It uses the functions defined
in Game and NeuralNet. args are specified in main.py.
"""
def __init__(self, game, nnet, args):
self.game = game
self.nnet = nnet
self.pnet = self.nnet.__class__(self.game) # the competitor network
self.args = args
self.mcts = MCTS(self.game, self.nnet, self.args)
self.trainExamplesHistory = [] # history of examples from args.numItersForTrainExamplesHistory latest iterations
self.skipFirstSelfPlay = False # can be overriden in loadTrainExamples()
def executeEpisode(self):
"""
This function executes one episode of self-play, starting with player 1.
As the game is played, each turn is added as a training example to
trainExamples. The game is played till the game ends. After the game
ends, the outcome of the game is used to assign values to each example
in trainExamples.
It uses a temp=1 if episodeStep < tempThreshold, and thereafter
uses temp=0.
Returns:
trainExamples: a list of examples of the form (canonicalBoard,pi,v)
pi is the MCTS informed policy vector, v is +1 if
the player eventually won the game, else -1.
"""
trainExamples = []
board = self.game.getInitBoard()
self.curPlayer = 1
episodeStep = 0
while True and episodeStep<200:
episodeStep += 1
canonicalBoard = self.game.getCanonicalForm(board,self.curPlayer)
valids = self.game.getValidMoves(canonicalBoard, 1)
temp = int(episodeStep < self.args.tempThreshold)
pi = self.mcts.getActionProb(canonicalBoard, temp=temp)
if np.sum(pi) == 0: break
#sym = self.game.getSymmetries(canonicalBoard, pi)
#for b,p in sym:
# trainExamples.append([b, self.curPlayer, p, None])
#self.game.print_board(canonicalBoard)
action = np.random.choice(len(pi), p=pi)
trainExamples.append([canonicalBoard, self.curPlayer, pi, None, valids, episodeStep])
board, self.curPlayer = self.game.getNextState(board, self.curPlayer, action)
r = self.game.getGameEnded(board, self.curPlayer)
if r!=0:
return [(x[0],x[2],r*x[1], x[4], x[5], episodeStep) for x in trainExamples]
#return [(x[0],x[2],0) for x in trainExamples]
print("the game's not ended")
return []
def learn(self):
"""
Performs numIters iterations with numEps episodes of self-play in each
iteration. After every iteration, it retrains neural network with
examples in trainExamples (which has a maximium length of maxlenofQueue).
It then pits the new neural network against the old one and accepts it
only if it wins >= updateThreshold fraction of games.
"""
for i in range(1, self.args.numIters+1):
# bookkeeping
print('------ITER ' + str(i) + '------')
# examples of the iteration
if not self.skipFirstSelfPlay or i>1:
iterationTrainExamples = deque([], maxlen=self.args.maxlenOfQueue)
eps_time = AverageMeter()
bar = Bar('Self Play', max=self.args.numEps)
end = time.time()
for eps in range(self.args.numEps):
self.mcts = MCTS(self.game, self.nnet, self.args) # reset search tree
iterationTrainExamples += self.executeEpisode()
# bookkeeping + plot progress
eps_time.update(time.time() - end)
end = time.time()
bar.suffix = '({eps}/{maxeps}) Eps Time: {et:.3f}s | Total: {total:} | ETA: {eta:}'.format(eps=eps+1, maxeps=self.args.numEps, et=eps_time.avg,
total=bar.elapsed_td, eta=bar.eta_td)
bar.next()
bar.finish()
# save the iteration examples to the history
self.trainExamplesHistory.append(iterationTrainExamples)
trainStats = [0,0,0]
for res in iterationTrainExamples:
trainStats[res[2]] += 1
print(trainStats)
if len(self.trainExamplesHistory) > self.args.numItersForTrainExamplesHistory:
print("len(trainExamplesHistory) =", len(self.trainExamplesHistory), " => remove the oldest trainExamples")
self.trainExamplesHistory.pop(0)
# backup history to a file
# NB! the examples were collected using the model from the previous iteration, so (i-1)
self.saveTrainExamples(i-1)
# shuffle examlpes before training
trainExamples = []
for e in self.trainExamplesHistory:
trainExamples.extend(e)
shuffle(trainExamples)
# training new network, keeping a copy of the old one
self.nnet.save_checkpoint(folder=self.args.checkpoint, filename='temp.pth.tar')
self.pnet.load_checkpoint(folder=self.args.checkpoint, filename='temp.pth.tar')
pmcts = MCTS(self.game, self.pnet, self.args)
self.nnet.train(trainExamples)
nmcts = MCTS(self.game, self.nnet, self.args)
print('PITTING AGAINST PREVIOUS VERSION')
arena = Arena(lambda x: np.argmax(pmcts.getActionProb(x, temp=0)),
lambda x: np.argmax(nmcts.getActionProb(x, temp=0)), self.game)
pwins, nwins, draws = arena.playGames(self.args.arenaCompare)
print('NEW/PREV WINS : %d / %d ; DRAWS : %d' % (nwins, pwins, draws))
if pwins+nwins > 0 and float(nwins)/(pwins+nwins) < self.args.updateThreshold:
print('REJECTING NEW MODEL')
self.nnet.load_checkpoint(folder=self.args.checkpoint, filename='temp.pth.tar')
else:
print('ACCEPTING NEW MODEL')
self.nnet.save_checkpoint(folder=self.args.checkpoint, filename=self.getCheckpointFile(i))
self.nnet.save_checkpoint(folder=self.args.checkpoint, filename='best.pth.tar')
def getCheckpointFile(self, iteration):
return 'checkpoint_' + str(iteration) + '.pth.tar'
def saveTrainExamples(self, iteration):
folder = self.args.checkpoint
if not os.path.exists(folder):
os.makedirs(folder)
filename = os.path.join(folder, self.getCheckpointFile(iteration)+".examples")
with open(filename, "wb+") as f:
Pickler(f).dump(self.trainExamplesHistory)
f.closed
def loadTrainExamples(self):
modelFile = os.path.join(self.args.load_folder_examples_file[0], self.args.load_folder_examples_file[1])
examplesFile = modelFile+".examples"
if not os.path.isfile(examplesFile):
print(examplesFile)
r = input("File with trainExamples not found. Continue? [y|n]")
if r != "y":
sys.exit()
else:
print("File with trainExamples found. Read it.")
with open(examplesFile, "rb") as f:
self.trainExamplesHistory = Unpickler(f).load()
f.closed
# examples based on the model were already collected (loaded)
self.skipFirstSelfPlay = True
|