def main():
directory = f'./saves/{args.id}{"_early" if args.early_stop else ""}'
if not args.live and not args.test_live:
try: shutil.rmtree(directory)
except: pass
hs = HSearchEnv(gpu_split=args.gpu_split, net_type=args.net_type)
flat, hydrated, network = hs.get_winner(id=args.id)
env = BitcoinEnv(flat, name='ppo_agent')
agent = agents_dict['ppo_agent'](
saver_spec=dict(
directory=directory,
# saves this model every 6000 time-steps. I'd rather manually save it at the end, that way we could save
# a winning combo in hypersearch.py and remove this redundant training step - but TForce doesn't have
# working manual-save code yet, only automatic.
steps=6000
),
states_spec=env.states,
actions_spec=env.actions,
network_spec=network,
**hydrated
)
if args.live or args.test_live:
env.run_live(agent, test=args.test_live)
else:
env.train_and_test(agent, early_stop=args.early_stop, n_tests=args.runs)
agent.close()
env.close()
def main():
directory = os.path.join(os.getcwd(), "saves", args.name)
filestar = os.path.join(directory, args.name)
live_ish = args.live or args.test_live
if not live_ish:
try:
shutil.rmtree(directory)
except:
pass
os.mkdir(directory)
hs = HSearchEnv(cli_args=args)
flat, hydrated, network = hs.get_winner(id=args.id)
env = BitcoinEnv(flat, cli_args=args)
agent = agents_dict['ppo_agent'](states=env.states,
actions=env.actions,
network=network,
**hydrated)
if live_ish:
agent.restore_model(directory)
env.run_live(agent, test=args.test_live)
else:
env.train_and_test(agent, args.n_steps, args.n_tests, args.early_stop)
agent.save_model(filestar)
agent.close()
env.close()
def main():
directory = os.path.join(os.getcwd(), "saves", args.name)
filestar = os.path.join(directory, args.name)
live_ish = args.live or args.test_live
if not live_ish:
try: shutil.rmtree(directory)
except: pass
os.mkdir(directory)
hs = HSearchEnv(cli_args=args)
flat, hydrated, network = hs.get_winner(id=args.id)
env = BitcoinEnv(flat, name='ppo_agent')
agent = agents_dict['ppo_agent'](
states_spec=env.states,
actions_spec=env.actions,
network_spec=network,
**hydrated
)
if live_ish:
agent.restore_model(directory)
env.run_live(agent, test=args.test_live)
else:
env.train_and_test(agent, args.n_steps, args.n_tests, args.early_stop)
agent.save_model(filestar)
agent.close()
env.close()