def test_agent_online(robot, game, model, budget):
# to address game -> we have a new game here
i = 0
queried_x = []
queried_y = []
performance = []
test_sents = helpers.data2sents(game.test_x, game.test_y)
game.reboot()
while i < budget:
sel_ind = game.current_frame
# construct the observation
observation = game.get_frame(model)
action = robot.get_action(observation)
if action[1] == 1:
sentence = game.train_x[sel_ind]
labels = game.train_y[sel_ind]
queried_x.append(sentence)
queried_y.append(labels)
i += 1
train_sents = helpers.data2sents(queried_x, queried_y)
model.train(train_sents)
performance.append(model.test(test_sents))
reward, observation2, terminal = game.feedback(action, model) # game
robot.update(observation, action, reward, observation2, terminal)
# train a crf and evaluate it
train_sents = helpers.data2sents(queried_x, queried_y)
model.train(train_sents)
performance.append(model.test(test_sents))
print("***TEST", performance)
def test_agent_online(robot, game, model, budget):
# to address game -> we have a new game here
i = 0
queried_x = []
queried_y = []
performance = []
test_sents = helpers.data2sents(game.test_x, game.test_y)
X_test = [tagger.sent2features(s) for s in test_sents]
Y_true = [tagger.sent2labels(s) for s in test_sents]
game.reboot(model)
while i < budget:
sel_ind = game.current_frame
# construct the observation
observation = game.get_frame(model)
action = robot.get_action(observation)
if action[1] == 1:
sentence = game.train_x[sel_ind]
labels = game.train_y[sel_ind]
queried_x.append(sentence)
queried_y.append(labels)
i += 1
train_sents = helpers.data2sents(queried_x, queried_y)
model.train(train_sents)
performance.append(model.test(X_test, Y_true))
reward, observation2, terminal = game.feedback(action, model) # game
robot.update(observation, action, reward, observation2, terminal)
# train a crf and evaluate it
train_sents = helpers.data2sents(queried_x, queried_y)
model.train(train_sents)
performance.append(model.test(X_test, Y_true))
logging.debug('***TEST {}'.format(performance))
def test_agent_batch(robot, game, model, budget):
i = 0
queried_x = []
queried_y = []
performance = []
cost = []
test_sents = helpers.data2sents(game.test_x, game.test_y)
game.replay()
while i < budget:
sel_ind = game.current_frame
# construct the observation
observation = game.get_frame(model)
action = robot.get_action(observation)
if action[1] == 1:
sentence = game.train_x[sel_ind]
labels = game.train_y[sel_ind]
queried_x.append(sentence)
queried_y.append(labels)
i += 1
train_sents = helpers.data2sents(queried_x, queried_y)
model.train(train_sents)
performance.append(model.test(test_sents))
cost.append(i)
game.current_frame += 1
print('***COST', cost)
print("***TEST", performance)
return (cost, performance)
def test_agent_batch(robot, game, model, budget):
i = 0
queried_x = []
queried_y = []
performance = []
test_sents = helpers.data2sents(game.test_x, game.test_y)
X_test = [tagger.sent2features(s) for s in test_sents]
Y_true = [tagger.sent2labels(s) for s in test_sents]
game.reboot(model)
while i < budget:
sel_ind = game.current_frame
# construct the observation
observation = game.get_frame(model)
action = robot.get_action(observation)
if action[1] == 1:
sentence = game.train_x[sel_ind]
labels = game.train_y[sel_ind]
queried_x.append(sentence)
queried_y.append(labels)
i += 1
train_sents = helpers.data2sents(queried_x, queried_y)
model.train(train_sents)
performance.append(model.test(X_test, Y_true))
game.current_frame += 1
# train a crf and evaluate it
train_sents = helpers.data2sents(queried_x, queried_y)
model.train(train_sents)
performance.append(model.test(X_test, Y_true))
logging.debug('***TEST {}'.format(performance))
def test_agent_batch(robot, game, model, budget):
i = 0
queried_x = []
queried_y = []
performance = []
test_sents = helpers.data2sents(game.test_x, game.test_y)
game.reboot()
while i < budget:
sel_ind = game.current_frame
# construct the observation
observation = game.get_frame(model)
action = robot.get_action(observation)
if action[1] == 1:
sentence = game.train_x[sel_ind]
labels = game.train_y[sel_ind]
queried_x.append(sentence)
queried_y.append(labels)
i += 1
train_sents = helpers.data2sents(queried_x, queried_y)
model.train(train_sents)
performance.append(model.test(test_sents))
game.current_frame += 1
# train a crf and evaluate it
train_sents = helpers.data2sents(queried_x, queried_y)
model.train(train_sents)
performance.append(model.test(test_sents))
print "***TEST", performance
def get_performance(self, tagger):
# train with {queried_set_x, queried_set_y}
# train with examples: self.model.train(self.queried_set_x,
# self.queried_set_y)
if tagger.name == "RNN":
tagger.train(self.queried_set_idx, self.queried_set_y)
performance = tagger.test(self.dev_idx, self.dev_y)
return performance
print (len(self.queried_set_x), len(self.queried_set_y))
train_sents = helpers.data2sents(
self.queried_set_x, self.queried_set_y)
# print train_sents
tagger.train(train_sents)
# test on development data
test_sents = helpers.data2sents(self.dev_x, self.dev_y)
performance = tagger.test(test_sents)
#performance = self.model.test2conlleval(self.dev_x, self.dev_y)
return performance
def get_performance(self, tagger):
# train with {queried_set_x, queried_set_y}
# train with examples: self.model.train(self.queried_set_x, self.queried_set_y)
if tagger.name == 'RNN':
tagger.train(self.queried_set_idx, self.queried_set_idy)
performance = tagger.test(features=self.dev_idx, labels=self.dev_y)
return performance
else:
# CRF case
train_sents = helpers.data2sents(self.queried_set_x,
self.queried_set_y)
tagger.train(train_sents)
# test on development data
performance = tagger.test(self.X_dev, self.Y_dev)
# performance = self.model.test2conlleval(self.dev_x, self.dev_y)
return performance
def __init__(self, story, dev, max_len, w2v, budget, model_name):
# build environment
# load data as story
logging.info('Initializing the game:')
# import story
self.train_x, self.train_y, self.train_idx, self.train_idy = story
self.dev_x, self.dev_y, self.dev_idx, self.dev_idy = dev
# TODO this should be a function
# Cache only for CRF model
if model_name == 'CRF':
dev_sents = helpers.data2sents(self.dev_x, self.dev_y)
self.X_dev = [tagger.sent2features(s) for s in dev_sents]
self.Y_dev = [tagger.sent2labels(s) for s in dev_sents]
self.max_len = max_len
self.w2v = w2v
logging.info('Story: length = {}'.format(len(self.train_x)))
self.order = list(range(0, len(self.train_x)))
# if re-order, use random.shuffle(self.order)
# load word embeddings, pretrained - w2v
logging.info('Dictionary size {} Embedding size {}'.format(
len(self.w2v), len(self.w2v[0])))
# when queried times is 'budget', then stop
self.budget = budget
self.queried_times = 0
# TODO use ndarrays for everything
# select pool
self.queried_set_x = []
self.queried_set_y = []
self.queried_set_idx = []
self.queried_set_idy = []
# let's start
self.episode = 0
# story frame
self.current_frame = 0
#self.nextFrame = self.current_frame + 1
self.terminal = False
self.make_query = False
self.performance = 0