def train(cls, num_frames: int):
rl_agent = cls()
preset = ddpg(device=rl_agent.device,
lr_q=Settings.LEARNING_RATE,
lr_pi=Settings.LEARNING_RATE)
experiment = SingleEnvExperiment(preset, rl_agent.env)
experiment.train(num_frames)
default_log_dir = experiment._writer.log_dir
copy_tree(default_log_dir, Settings.FULL_LOG_DIR)
rmtree(default_log_dir)
rl_agent.env.close()
def resume_training(cls, path, num_frames: int):
rl_agent = cls()
lr = Settings.LEARNING_RATE
agent = rainbow(device=rl_agent.device, lr=lr)
q_dist_module = torch.load(os.path.join(path, "q_dist.pt"), map_location='cpu').to(rl_agent.device)
experiment = SingleEnvExperiment(agent, rl_agent.env)
agent = experiment._agent
old_q_dist = agent.q_dist
old_q_dist.model.load_state_dict(q_dist_module.state_dict())
experiment.train(frames=num_frames)
default_log_dir = experiment._writer.log_dir
copy_tree(default_log_dir, Settings.FULL_LOG_DIR)
rmtree(default_log_dir)
rl_agent.env.close()
def train_dqn_all():
from all.environments import GymEnvironment
from all.presets.classic_control import ddqn
from all.experiments import SingleEnvExperiment
if Settings.CUDA:
device = "cuda"
else:
device = "cpu"
env = GymEnvironment(Settings.GYM_ENVIRONMENT, device=device)
preset = ddqn(device=device,
lr=Settings.LEARNING_RATE,
initial_exploration=Settings.EPS_START,
final_exploration=Settings.EPS_END)
experiment = SingleEnvExperiment(preset, env)
experiment.train(1E6)
default_log_dir = experiment._writer.log_dir
copy_tree(default_log_dir, Settings.FULL_LOG_DIR)
rmtree(default_log_dir)
def resume_dqn_all():
from all.presets.classic_control import ddqn
from all.environments import GymEnvironment
from all.experiments import SingleEnvExperiment
if Settings.CUDA:
device = "cuda"
else:
device = "cpu"
env = GymEnvironment('sumo-jerk-v0', device=device)
lr = 1e-5
agent = ddqn(device=device, lr=lr)
q_module = torch.load(os.path.join('models', "q.pt"),
map_location='cpu').to(device)
experiment = SingleEnvExperiment(agent, env)
agent = experiment._agent
old_q = agent.q
old_q.model.load_state_dict(q_module.state_dict())
experiment.train(frames=1e6)
default_log_dir = experiment._writer.log_dir
copy_tree(default_log_dir, Settings.FULL_LOG_DIR)
rmtree(default_log_dir)
previous_runs = set(get_existing_runs())
env = gym.make('nextro-v0', c_args=args)
env = GymEnvironment(env, args.device)
pretrained_models = None
if args.loc != '':
pretrained_models = load_weights(args.loc)
agent = sac_minitaur_inspired(device=args.device,
last_frame=args.frames,
pretrained_models=pretrained_models,
train_parallel=args.train_parallel)
exp = SingleEnvExperiment(agent, env, render=args.render)
all_runs = set(get_existing_runs())
current_run_name = list(all_runs - previous_runs)[0]
if args.mode == 'train':
if args.episodes != 0:
exp.train(episodes=args.episodes)
else:
exp.train(frames=args.frames)
else:
if args.episodes == 0:
raise Exception('You HAVE TO speficy EPISODES when using test!!!')
exp.test(episodes=args.episodes)
exp._env._env.store_settings(current_run_name)