def play_ner(replay, saved_robot, saved_initdata, nextepisode):
actions = 2
global AGENT
if AGENT == "random":
robot = RobotRandom(actions)
elif AGENT == "DQN":
robot = RobotDQN(actions)
elif AGENT == "CNNDQN":
robot = RobotCNNDQN(actions)
else:
print "** There is no robot."
raise SystemExit
if replay:
robot.restore(saved_robot)
global TRAIN_LANG, TRAIN_LANG_NUM, BUDGET, CHECKPOINT
robot.setCheckPoint(CHECKPOINT)
for i in range(TRAIN_LANG_NUM):
train = TRAIN_LANG[i][0]
test = TRAIN_LANG[i][1]
dev = TRAIN_LANG[i][2]
emb = TRAIN_LANG[i][3]
tagger = TRAIN_LANG[i][4]
# initilise a NER game
game = initialise_game(train, test, dev, emb, BUDGET, True, replay, saved_initdata)
# initialise a decision robot
robot.update_embeddings(game.w2v)
# tagger
model = CRFTagger(tagger)
model.clean()
# play game
episode = 1
if replay:
episode = nextepisode
print(">>>>>> Playing game ..")
while episode <= MAX_EPISODE:
#copy the baseline model to the saved model
#print tagger
#shutil.copy('eng.model.baseline', tagger)
print '>>>>>>> Current game round ', episode, 'Maximum ', MAX_EPISODE
observation = game.get_frame(model)
action = robot.get_action(observation)
print '> Action', action
reward, observation2, terminal = game.feedback(action, model, True)
print '> Reward', reward
robot.update(observation, action, reward, observation2, terminal)
if terminal == True:
episode += 1
print '> Terminal <'
return robot
def test(robot):
global TEST_LANG, TEST_LANG_NUM, BUDGET
for i in range(TEST_LANG_NUM):
train = TEST_LANG[i][0]
test = TEST_LANG[i][1]
dev = TEST_LANG[i][2]
emb = TEST_LANG[i][3]
tagger = TEST_LANG[i][4]
game2 = initialise_game(train, test, dev, emb, BUDGET, False, False, '')
robot.update_embeddings(game2.w2v)
model = CRFTagger(tagger)
model.clean()
test_agent_batchNew(robot, game2, model, 1000, True)