def start_session(self):
""" Launch the tensorflow session and start the GuessWhat loop """
with tf.Session(config=self.tf_config) as sess:
guesser_network = GuesserNetwork(self.guesser_config['model'],
num_words=self.tokenizer.no_words)
guesser_var = [
v for v in tf.global_variables() if 'guesser' in v.name
]
guesser_saver = tf.train.Saver(var_list=guesser_var)
guesser_saver.restore(sess, GUESS_NTW_PATH)
guesser_wrapper = GuesserROSWrapper(guesser_network)
qgen_network = QGenNetworkLSTM(self.qgen_config['model'],
num_words=self.tokenizer.no_words,
policy_gradient=False)
qgen_var = [v for v in tf.global_variables() if 'qgen' in v.name]
qgen_saver = tf.train.Saver(var_list=qgen_var)
qgen_saver.restore(sess, QGEN_NTW_PATH)
qgen_network.build_sampling_graph(
self.qgen_config['model'],
tokenizer=self.tokenizer,
max_length=self.eval_config['loop']['max_depth'])
qgen_wrapper = QGenWrapper(
qgen_network,
self.tokenizer,
max_length=self.eval_config['loop']['max_depth'],
k_best=self.eval_config['loop']['beam_k_best'])
oracle_wrapper = OracleROSWrapper(self.tokenizer)
self.loop(sess, guesser_wrapper, qgen_wrapper, oracle_wrapper)
def __init__(self, config, oracle, qgen, guesser, tokenizer):
self.storage = []
self.tokenizer = tokenizer
self.batch_size = config["optimizer"]["batch_size"]
self.max_no_question = config['loop']['max_question']
self.max_depth = config['loop']['max_depth']
self.k_best = config['loop']['beam_k_best']
self.oracle = OracleWrapper(oracle, tokenizer)
self.guesser = GuesserWrapper(guesser)
self.qgen = QGenWrapper(qgen, tokenizer, max_length=self.max_depth, k_best=self.k_best)
oracle_saver.restore(sess, os.path.join(args.networks_dir, 'oracle', loop_config["oracle_identifier"], 'best', 'params.ckpt'))
guesser_saver.restore(sess, os.path.join(args.networks_dir, 'guesser', loop_config["guesser_identifier"], 'best', 'params.ckpt'))
# create training tools
loop_sources = qgen_network.get_sources(sess)
logger.info("Sources: " + ', '.join(loop_sources))
train_batchifier = LooperBatchifier(tokenizer, generate_new_games=True)
eval_batchifier = LooperBatchifier(tokenizer, generate_new_games=False)
# Initialize the looper to eval/train the game-simulation
qgen_batchifier = qgen_batchifier_cstor(tokenizer, sources=qgen_network.get_sources(sess), generate=True)
qgen_wrapper = QGenWrapper(qgen_network, qgen_batchifier, tokenizer,
max_length=loop_config['loop']['max_depth'],
k_best=loop_config['loop']['beam_k_best'])
oracle_split_mode = 1
# oracle_split_mode = BatchifierSplitMode.from_string(oracle_config["model"]["question"]["input_type"])
oracle_batchifier = oracle_batchifier_cstor(tokenizer, sources=oracle_network.get_sources(sess), split_mode=oracle_split_mode)
oracle_wrapper = OracleWrapper(oracle_network, oracle_batchifier, tokenizer)
guesser_batchifier = guesser_batchifier_cstor(tokenizer, sources=guesser_network.get_sources(sess))
guesser_wrapper = GuesserWrapper(guesser_network, guesser_batchifier, tokenizer, guesser_listener)
xp_manager.configure_score_tracking("valid_accuracy", max_is_best=True)
game_engine = BasicLooper(loop_config,
oracle_wrapper=oracle_wrapper,
guesser_wrapper=guesser_wrapper,
qgen_wrapper=qgen_wrapper,
split_mode=oracle_split_mode)
oracle_wrapper = OracleWrapper(oracle_network, oracle_batchifier,
tokenizer)
guesser_batchifier = guesser_batchifier_cstor(
tokenizer, sources=guesser_network.get_sources(sess_loop))
guesser_wrapper = GuesserWrapper(guesser_network, guesser_batchifier,
tokenizer, guesser_listener)
qgen_batchifier = qgen_batchifier_cstor(
tokenizer,
sources=qgen_network.get_sources(sess_loop),
generate=True)
qgen_wrapper = QGenWrapper(qgen_network,
qgen_batchifier,
tokenizer,
max_length=12,
k_best=20)
xp_manager.configure_score_tracking("valid_accuracy", max_is_best=True)
loop_config = {} # fake config
loop_config['loop'] = {}
loop_config['loop']['max_question'] = 5
game_engine = BasicLooper(loop_config,
oracle_wrapper=oracle_wrapper,
guesser_wrapper=guesser_wrapper,
qgen_wrapper=qgen_wrapper,
tokenizer=tokenizer,
batch_size=64)
qgen_var = [v for v in tf.global_variables()
if "qgen" in v.name] # and 'rl_baseline' not in v.name
qgen_saver = tf.train.Saver(var_list=qgen_var)
qgen_saver.restore(
sess,
os.path.join(args.networks_dir, 'qgen', args.qgen_identifier,
'params.ckpt'))
qgen_network.build_sampling_graph(
qgen_config["model"],
tokenizer=tokenizer,
max_length=eval_config['loop']['max_depth'])
qgen_wrapper = QGenWrapper(
qgen_network,
tokenizer,
max_length=eval_config['loop']['max_depth'],
k_best=eval_config['loop']['beam_k_best'])
else:
qgen_wrapper = QGenUserWrapper(tokenizer)
logger.info("No QGen was registered >>> use user input")
looper_evaluator = BasicLooper(eval_config,
oracle_wrapper=oracle_wrapper,
guesser_wrapper=guesser_wrapper,
qgen_wrapper=qgen_wrapper,
tokenizer=tokenizer,
batch_size=1)
logs = []