config.lwf = args.lwf
config.lwf_loss = args.lwf_loss
config.lwf_weight = args.lwf_weight
config.num_old_actions = int(args.num_old_actions)
return config
def print_config(config):
print 'Current config:\n'
variables = zip(vars(config).keys(), vars(config).values())
for var, val in sorted(variables):
print var + ' = ' + str(val)
if __name__ == '__main__':
args = parse_args()
my_config = modify_config(args)
print_config(my_config)
with tf.device('/gpu:' + str(args.gpu)):
# make env
env = gym.make(my_config.env_name)
env = wrap_dqn(env)
model = NatureQN(env, my_config)
model.initialize_eval()
model.evaluate()
if my_config.record:
model.record()
of the training is to use Tensorboard. The starter code writes summaries of different
variables.
To launch tensorboard, open a Terminal window and run
tensorboard --logdir=results/
Then, connect remotely to
address-ip-of-the-server:6006
6006 is the default port used by tensorboard.
"""
if __name__ == '__main__':
# make env
starttime = time.time()
env = gym.make(config.env_name)
env = MaxAndSkipEnv(env, skip=config.skip_frame)
env = PreproWrapper(env,
prepro=greyscale,
shape=(80, 80, 1),
overwrite_render=config.overwrite_render)
# exploration strategy
exp_schedule = LinearExploration(env, config.eps_begin, config.eps_end,
config.eps_nsteps)
# learning rate schedule
lr_schedule = LinearSchedule(config.lr_begin, config.lr_end,
config.lr_nsteps)
# train model
model = NatureQN(env, config)
model.run(exp_schedule, lr_schedule)
print('Total render time : {:.2f}'.format(time.time() - starttime))
of the training is to use Tensorboard. The starter code writes summaries of different
variables.
To launch tensorboard, open a Terminal window and run
tensorboard --logdir=results/
Then, connect remotely to
address-ip-of-the-server:6006
6006 is the default port used by tensorboard.
"""
if __name__ == '__main__':
# make env
env = gym.make(config.env_name)
env = MaxAndSkipEnv(env, skip=config.skip_frame)
env = PreproWrapper(env, prepro=greyscale, shape=(80, 80, 1),
overwrite_render=config.overwrite_render)
# exploration strategy
exp_schedule = LinearExploration(env, config.eps_begin,
config.eps_end, config.eps_nsteps)
# learning rate schedule
lr_schedule = LinearSchedule(config.lr_begin, config.lr_end,
config.lr_nsteps)
# train model
model = NatureQN(env, config)
model.original_schedule = config.original_schedule
model.logger.info('original schedule: {}'.format(model.original_schedule))
model.run(exp_schedule, lr_schedule)
env = PreproWrapper(env, prepro=greyscale, shape=(80, 80, 1),
overwrite_render=config.overwrite_render)
rewards = []
experts_meta_lis = [
'./core/checkpoints/q_learning/skip_connection/q5_train_atari_nature/deepdqn_weights/.meta', './core/checkpoints/q_learning/skip_connection/q5_train_atari_nature/resnet_weights/.meta', './core/checkpoints/policy_gradients/policy_network.ckpt.meta']
experts_chkpt_lis = [
'./core/checkpoints/q_learning/skip_connection/q5_train_atari_nature/deepdqn_weights/', './core/checkpoints/q_learning/skip_connection/q5_train_atari_nature/resnet_weights/', './core/checkpoints/policy_gradients/policy_network.ckpt']
experts = []
#temp_sess = None
for meta_path, chkpt_path in zip(experts_meta_lis, experts_chkpt_lis):
print([n.name for n in tf.get_default_graph().as_graph_def().node])
if "deepdqn" in meta_path:
model = NatureQN(env, config)
if "resnet" in meta_path:
model = ResnetQN(env, config)
if "policy" in meta_path:
continue
# if temp_sess == None:
#temp_sess = model.sess
model.initialize(meta_path, chkpt_path)
experts.append(model)
# with model.graph.as_default():
print("LOADED ALL MODELS")
for i in range(len(experts)):
guide = experts[i]
guide_experience = [[]]
If so, please report your hyperparameters.
You'll find the results, log and video recordings of your agent every 250k under
the corresponding file in the results folder. A good way to monitor the progress
of the training is to use Tensorboard. The starter code writes summaries of different
variables.
To launch tensorboard, open a Terminal window and run
tensorboard --logdir=results/
Then, connect remotely to
address-ip-of-the-server:6006
6006 is the default port used by tensorboard.
"""
if __name__ == '__main__':
# make env
env = gym.make(config.env_name)
env = MaxAndSkipEnv(env, skip=config.skip_frame)
env = PreproWrapper(env,
prepro=greyscale,
shape=(80, 80, 1),
overwrite_render=config.overwrite_render)
# load model
model = NatureQN(env, config)
model.initialize()
loaded = load_model(model)
assert loaded != False, "Loading failed"
# evaluate one episode of data
model.evaluate(env, 1)