def train_w_parameters(cmd):
args = parse_command(cmd)
# convert to dictionary
params = vars(args)
params['double_q'] = True
params['num_agent_train_steps_per_iter'] = 1
params['num_critic_updates_per_agent_update'] = 1
params['exploit_weight_schedule'] = ConstantSchedule(1.0)
params['video_log_freq'] = -1 # This param is not used for DQN
params['num_timesteps'] = 50000
params['learning_starts'] = 2000
params['eps'] = 0.2
##################################
### CREATE DIRECTORY FOR LOGGING
##################################
if params['env_name'] == 'PointmassEasy-v0':
params['ep_len'] = 50
if params['env_name'] == 'PointmassMedium-v0':
params['ep_len'] = 150
if params['env_name'] == 'PointmassHard-v0':
params['ep_len'] = 100
if params['env_name'] == 'PointmassVeryHard-v0':
params['ep_len'] = 200
if params['use_rnd']:
params['explore_weight_schedule'] = PiecewiseSchedule(
[(0, 1), (params['num_exploration_steps'], 0)], outside_value=0.0)
else:
params['explore_weight_schedule'] = ConstantSchedule(0.0)
if params['unsupervised_exploration']:
params['explore_weight_schedule'] = ConstantSchedule(1.0)
params['exploit_weight_schedule'] = ConstantSchedule(0.0)
if not params['use_rnd']:
params['learning_starts'] = params['num_exploration_steps']
logdir_prefix = 'hw5_expl_' # keep for autograder
data_path = os.path.join(os.path.dirname(os.path.realpath(__file__)),
'../../data')
if not (os.path.exists(data_path)):
os.makedirs(data_path)
logdir = logdir_prefix + args.exp_name + '_' + args.env_name + '_' + time.strftime(
"%d-%m-%Y_%H-%M-%S")
logdir = os.path.join(data_path, logdir)
params['logdir'] = logdir
if not (os.path.exists(logdir)):
os.makedirs(logdir)
print("\n\n\nLOGGING TO: ", logdir, "\n\n\n")
trainer = Q_Trainer(params)
trainer.run_training_loop()
def main():
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--env_name',
default='PointmassHard-v0',
choices=('PointmassEasy-v0', 'PointmassMedium-v0',
'PointmassHard-v0', 'PointmassVeryHard-v0'))
parser.add_argument('--use_modified_rew', action='store_true')
parser.add_argument('--exp_name', type=str, default='todo')
parser.add_argument('--eval_batch_size', type=int, default=1000)
parser.add_argument('--batch_size', type=int, default=256)
parser.add_argument('--use_rnd', action='store_true')
parser.add_argument('--num_exploration_steps', type=int, default=10000)
parser.add_argument('--unsupervised_exploration', action='store_true')
parser.add_argument('--offline_exploitation', action='store_true')
parser.add_argument('--cql_alpha', type=float, default=0.0)
parser.add_argument('--exploit_rew_shift', type=float, default=0.0)
parser.add_argument('--exploit_rew_scale', type=float, default=1.0)
parser.add_argument('--rnd_output_size', type=int, default=5)
parser.add_argument('--rnd_n_layers', type=int, default=2)
parser.add_argument('--rnd_size', type=int, default=400)
parser.add_argument('--seed', type=int, default=2)
parser.add_argument('--no_gpu', '-ngpu', action='store_true')
parser.add_argument('--which_gpu', '-gpu_id', default=0)
parser.add_argument('--scalar_log_freq', type=int, default=int(1e3))
parser.add_argument('--save_params', action='store_true')
args = parser.parse_args()
# convert to dictionary
params = vars(args)
params['double_q'] = True
params['num_agent_train_steps_per_iter'] = 1
params['num_critic_updates_per_agent_update'] = 1
params['exploit_weight_schedule'] = ConstantSchedule(1.0)
params['video_log_freq'] = -1 # This param is not used for DQN
params['num_timesteps'] = 50000
params['learning_starts'] = 2000
params['eps'] = 0.2
##################################
### CREATE DIRECTORY FOR LOGGING
##################################
if params['env_name'] == 'PointmassEasy-v0':
params['ep_len'] = 50
if params['env_name'] == 'PointmassMedium-v0':
params['ep_len'] = 150
if params['env_name'] == 'PointmassHard-v0':
params['ep_len'] = 100
if params['env_name'] == 'PointmassVeryHard-v0':
params['ep_len'] = 200
if params['use_rnd']:
params['explore_weight_schedule'] = PiecewiseSchedule(
[(0, 1), (params['num_exploration_steps'], 0)], outside_value=0.0)
else:
params['explore_weight_schedule'] = ConstantSchedule(0.0)
if params['unsupervised_exploration']:
params['explore_weight_schedule'] = ConstantSchedule(1.0)
params['exploit_weight_schedule'] = ConstantSchedule(0.0)
if not params['use_rnd']:
params['learning_starts'] = params['num_exploration_steps']
logdir_prefix = 'hw5_expl_' # keep for autograder
data_path = os.path.join(os.path.dirname(os.path.realpath(__file__)),
'../../data')
if not (os.path.exists(data_path)):
os.makedirs(data_path)
logdir = logdir_prefix + args.exp_name + '_' + args.env_name + '_' + time.strftime(
"%d-%m-%Y_%H-%M-%S")
logdir = os.path.join(data_path, logdir)
params['logdir'] = logdir
if not (os.path.exists(logdir)):
os.makedirs(logdir)
print("\n\n\nLOGGING TO: ", logdir, "\n\n\n")
trainer = Q_Trainer(params)
trainer.run_training_loop()
def main():
import argparse
parser = argparse.ArgumentParser()
parser.add_argument(
"--env_name",
default="PointmassHard-v0",
choices=(
"PointmassEasy-v0",
"PointmassMedium-v0",
"PointmassHard-v0",
"PointmassVeryHard-v0",
),
)
parser.add_argument("--exp_name", type=str, default="todo")
parser.add_argument("--eval_batch_size", type=int, default=1000)
parser.add_argument("--batch_size", type=int, default=256)
parser.add_argument("--use_rnd", action="store_true")
parser.add_argument("--num_exploration_steps", type=int, default=10000)
parser.add_argument("--unsupervised_exploration", action="store_true")
parser.add_argument("--offline_exploitation", action="store_true")
parser.add_argument("--cql_alpha", type=float, default=0.0)
parser.add_argument("--exploit_rew_shift", type=float, default=0.0)
parser.add_argument("--exploit_rew_scale", type=float, default=1.0)
parser.add_argument("--rnd_output_size", type=int, default=5)
parser.add_argument("--rnd_n_layers", type=int, default=2)
parser.add_argument("--rnd_size", type=int, default=400)
parser.add_argument("--seed", type=int, default=2)
parser.add_argument("--no_gpu", "-ngpu", action="store_true")
parser.add_argument("--which_gpu", "-gpu_id", default=0)
parser.add_argument("--scalar_log_freq", type=int, default=int(1e3))
parser.add_argument("--save_params", action="store_true")
# HW5: Custom exploration strategy
parser.add_argument("--modified_eps_greedy", action="store_true")
args = parser.parse_args()
# convert to dictionary
params = vars(args)
params["double_q"] = True
params["num_agent_train_steps_per_iter"] = 1
params["num_critic_updates_per_agent_update"] = 1
params["exploit_weight_schedule"] = ConstantSchedule(1.0)
params["video_log_freq"] = -1 # This param is not used for DQN
params["num_timesteps"] = 50000
params["learning_starts"] = 2000
params["eps"] = 0.2
##################################
### CREATE DIRECTORY FOR LOGGING
##################################
if params["env_name"] == "PointmassEasy-v0":
params["ep_len"] = 50
if params["env_name"] == "PointmassMedium-v0":
params["ep_len"] = 150
if params["env_name"] == "PointmassHard-v0":
params["ep_len"] = 100
if params["env_name"] == "PointmassVeryHard-v0":
params["ep_len"] = 200
if params["use_rnd"]:
params["explore_weight_schedule"] = PiecewiseSchedule(
[(0, 1), (params["num_exploration_steps"], 0)], outside_value=0.0
)
else:
params["explore_weight_schedule"] = ConstantSchedule(0.0)
if params["unsupervised_exploration"]:
params["explore_weight_schedule"] = ConstantSchedule(1.0)
params["exploit_weight_schedule"] = ConstantSchedule(0.0)
if not params["use_rnd"]:
params["learning_starts"] = params["num_exploration_steps"]
logdir_prefix = "hw5_expl_" # keep for autograder
data_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "../../data")
if not (os.path.exists(data_path)):
os.makedirs(data_path)
logdir = (
logdir_prefix
+ args.exp_name
+ "_"
+ args.env_name
+ "_"
+ time.strftime("%d-%m-%Y_%H-%M-%S")
)
logdir = os.path.join(data_path, logdir)
params["logdir"] = logdir
if not (os.path.exists(logdir)):
os.makedirs(logdir)
print("\n\n\nLOGGING TO: ", logdir, "\n\n\n")
trainer = Q_Trainer(params)
trainer.run_training_loop()