def _trigger_epoch(self):
if FEATURE:
if self.epoch_num % 1 == 0:
logger.info("update density network at epoch %d." %
(self.epoch_num))
cfg = PredictConfig(session_init=JustCurrentSession(),
model=Model(),
input_var_names=['state'],
output_var_names=[FEATURE])
self.offline_predictor = get_predict_func(cfg)
def run_atari_neptune_experiment(yamlFile=None, modelToLaod=None, epoch=None):
global ENV_NAME, EXPERIMENT_MODEL, FRAME_HISTORY
with open(yamlFile, 'r') as stream:
try:
yamlData = yaml.load(stream)
except yaml.YAMLError as exc:
print(exc)
argsDict = {}
for v in yamlData["parameters"]:
argsDict[v["name"]] = v["default"]
args = edict(argsDict)
ENV_NAME = args.env
assert ENV_NAME
if hasattr(args, "frame_history"):
FRAME_HISTORY = args.frame_history
# examples.OpenAIGym.train_atari_with_neptune.FRAME_HISTORY = args.frame_history
else:
FRAME_HISTORY = 4
# FRAME_HISTORY = int(get_atribute(args, "frame_history", 4))
logger.info("Environment Name: {}".format(ENV_NAME))
# module_name, function_name = ctx.params.featureExtractor.split(".")
module_name = args.experimentModelClass[:args.experimentModelClass.
rfind('.')]
class_name = args.experimentModelClass[args.experimentModelClass.
rfind('.') + 1:]
experiment_model_class = importlib.import_module(
module_name).__dict__[class_name]
EXPERIMENT_MODEL = experiment_model_class(args.experimentModelParameters)
p = get_player()
del p # set NUM_ACTIONS. Bloody hack!
EXPERIMENT_MODEL.set_number_of_actions(NUM_ACTIONS)
if args.gpu:
os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)
cfg = PredictConfig(model=EXPERIMENT_MODEL,
session_init=SaverRestore(modelToLaod),
input_var_names=['state'],
output_var_names=['logits'])
dump_dir = os.path.join(dump_dir_root, str(epoch))
print "Writing to:{}".format(dump_dir)
run_submission(cfg, dump_dir)
def _setup_graph(self):
self.sess = self.trainer.sess
self.async_predictor = MultiThreadAsyncPredictor(
self.trainer.get_predict_funcs(['state'], ['logitsT', 'pred_value'],
PREDICTOR_THREAD), batch_size=15)
# else:
# self.async_predictor = MultiThreadAsyncPredictor(
# self.trainer.get_predict_funcs(['state'], ['logitsT', 'pred_value', FEATURE],
# PREDICTOR_THREAD), batch_size=15)
if FEATURE:
logger.info("Initialize density network")
cfg = PredictConfig(
session_init=NewSession(),
model=Model(),
input_var_names=['state'],
output_var_names=[FEATURE])
self.offline_predictor = get_predict_func(cfg)
self.async_predictor.run()
"Conv2DBackpropFilter": "MKL",
"Conv2DBackpropInput": "MKL"
}
else:
print "using tensorflow convolution"
label_map = {}
with ops.Graph().as_default() as g:
tf.set_random_seed(my_task_index)
np.random.seed(my_task_index)
with g._kernel_label_map(label_map):
with tf.device('/job:worker/task:{}/cpu:0'.format(my_task_index)):
with tf.variable_scope(tf.get_variable_scope(), reuse=None):
if args.task != 'train':
cfg = PredictConfig(
model=Model('/job:ps/task:0/cpu:0'),
session_init=SaverRestore(args.load),
input_var_names=['state'],
output_var_names=['logits:0'])
if args.task == 'play':
play_model(cfg)
elif args.task == 'eval':
eval_model_multithread(cfg, EVAL_EPISODE)
else:
nr_towers = args.nr_towers
predict_towers = args.nr_predict_towers * [
0,
]
if args.cpu != 1:
nr_gpu = get_nr_gpu()
if nr_gpu > 1:
DUELING = False
else:
logger.error("dueling argument must be t or f")
if DOUBLE:
logger.info("Using Double")
if DUELING:
logger.info("Using Dueling")
assert ENV_NAME
logger.info("Environment Name: {}".format(ENV_NAME))
p = get_player()
del p # set NUM_ACTIONS
if args.gpu:
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
cfg = PredictConfig(model=Model(),
session_init=SaverRestore(args.load),
input_var_names=['state'],
output_var_names=['Qvalue'])
s_cfg = PredictConfig(model=Model(),
session_init=SaverRestore(args.load),
input_var_names=['state'],
output_var_names=['saliency'])
run(cfg, s_cfg, args.output)
#run_submission(cfg, args.output, args.episode)
#do_submit(args.output, args.api)
break
print float(det_cnt) / cnt
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--gpu', help='comma separated list of GPU(s) to use.')
parser.add_argument('--load', help='load model', required=True)
parser.add_argument('--env', help='environment name', required=True)
parser.add_argument('--split', help='split name', required=True)
args = parser.parse_args()
ENV_NAME = args.env
SPLIT_NAME = args.split
assert ENV_NAME
assert SPLIT_NAME
logger.info("Environment Name: {}".format(ENV_NAME))
logger.info("Split Name: {}".format(SPLIT_NAME))
p = get_player()
del p # set NUM_ACTIONS
if args.gpu:
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
cfg = PredictConfig(model=Model(),
session_init=SaverRestore(args.load),
input_names=['state'],
output_names=['logits', 'rnn_state'])
run_submission(cfg)
parser.add_argument('-o','--output', help='output directory', default='gym-submit')
parser.add_argument('-k','--key', help='api key')
args=parser.parse_args()
ENV_NAME = args.env
assert ENV_NAME
p = get_player(); del p # set NUM_ACTIONS
if args.gpu:
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
if args.task != 'train':
assert args.load is not None
if args.task != 'train':
cfg = PredictConfig(
model=Model(),
session_init=SaverRestore(args.load),
input_var_names=['state'],
output_var_names=['fct/output:0'])
if args.task == 'play':
play_model(cfg)
elif args.task == 'eval':
eval_model_multithread(cfg, EVAL_EPISODE)
elif args.task == 'run':
run_submission(cfg, args.output, args.episode)
do_submit(args.output, args.key)
else:
config = get_config()
if args.load:
config.session_init = SaverRestore(args.load)
QueueInputTrainer(config).train()