def __init__(self, args):
self.args = args
# seeding
utils.seed(args.seed)
self.env = gym.make(id=args.env)
self.episodes = 300 # args.episodes
self.horizon = self.env.max_steps
self.initial_decay = 0.99 # args.decay
self.observation_preprocessor = utils.ObssPreprocessor(
model_name=args.model,
obs_space=self.env.observation_space,
load_vocab_from=getattr(self.args, 'pretrained_model', None))
# TODO: for now I am only running the small model
self.model = models.ACModel(obs_space=self.env.observation_space,
action_space=self.env.action_space)
self.learner = ModelAgent(
model_or_name=self.model,
obss_preprocessor=self.observation_preprocessor,
argmax=True)
self.teacher = Bot(self.env)
self.data = []
self.observation_preprocessor.vocab.save()
utils.save_model(self.model, args.model)
self.model.train()
if torch.cuda.is_available():
self.model.cuda()
self.optimizer = torch.optim.Adam(self.model.parameters(),
self.args.lr,
eps=self.args.optim_eps)
self.scheduler = torch.optim.lr_scheduler.StepLR(self.optimizer,
step_size=100,
gamma=0.9)
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
if self.device.type == 'cpu':
print('running on cpu...')
if args.pretrained_model:
default_model_name = args.pretrained_model + '_pretrained_' + default_model_name
args.model = args.model.format(
**model_name_parts) if args.model else default_model_name
utils.configure_logging(args.model)
logger = logging.getLogger(__name__)
# Define obss preprocessor
if 'emb' in args.arch:
obss_preprocessor = utils.IntObssPreprocessor(args.model,
envs[0].observation_space,
args.pretrained_model)
else:
obss_preprocessor = utils.ObssPreprocessor(args.model,
envs[0].observation_space,
args.pretrained_model)
# Define actor-critic model
acmodel = utils.load_model(args.model, raise_not_found=False)
if acmodel is None:
if args.pretrained_model:
acmodel = utils.load_model(args.pretrained_model, raise_not_found=True)
else:
acmodel = ACModel(obss_preprocessor.obs_space,
envs[0].action_space,
args.image_dim,
args.memory_dim,
args.instr_dim,
not args.no_desc,
args.instr_arch,
def __init__(
self,
args,
):
self.args = args
utils.seed(self.args.seed)
# args.env is a list when training on multiple environments
if getattr(args, 'multi_env', None):
self.env = [gym.make(item) for item in args.multi_env]
self.train_demos = []
for demos, episodes in zip(args.multi_demos, args.multi_episodes):
demos_path = utils.get_demos_path(demos,
None,
None,
valid=False)
logger.info('loading {} of {} demos'.format(episodes, demos))
train_demos = utils.load_demos(demos_path)
logger.info('loaded demos')
if episodes > len(train_demos):
raise ValueError(
"there are only {} train demos in {}".format(
len(train_demos), demos))
self.train_demos.extend(train_demos[:episodes])
logger.info('So far, {} demos loaded'.format(
len(self.train_demos)))
self.val_demos = []
for demos, episodes in zip(args.multi_demos, [args.val_episodes] *
len(args.multi_demos)):
demos_path_valid = utils.get_demos_path(demos,
None,
None,
valid=True)
logger.info('loading {} of {} valid demos'.format(
episodes, demos))
valid_demos = utils.load_demos(demos_path_valid)
logger.info('loaded demos')
if episodes > len(valid_demos):
logger.info(
'Using all the available {} demos to evaluate valid. accuracy'
.format(len(valid_demos)))
self.val_demos.extend(valid_demos[:episodes])
logger.info('So far, {} valid demos loaded'.format(
len(self.val_demos)))
logger.info('Loaded all demos')
observation_space = self.env[0].observation_space
action_space = self.env[0].action_space
else:
self.env = gym.make(self.args.env)
demos_path = utils.get_demos_path(args.demos,
args.env,
args.demos_origin,
valid=False)
demos_path_valid = utils.get_demos_path(args.demos,
args.env,
args.demos_origin,
valid=True)
print("else")
logger.info('loading demos')
self.train_demos = utils.load_demos(demos_path)
print(len(self.train_demos))
print(self.train_demos[0])
logger.info('loaded demos')
if args.episodes:
if args.episodes > len(self.train_demos):
raise ValueError("there are only {} train demos".format(
len(self.train_demos)))
self.train_demos = self.train_demos[:args.episodes]
self.val_demos = utils.load_demos(demos_path_valid)
if args.val_episodes > len(self.val_demos):
logger.info(
'Using all the available {} demos to evaluate valid. accuracy'
.format(len(self.val_demos)))
self.val_demos = self.val_demos[:self.args.val_episodes]
observation_space = self.env.observation_space
action_space = self.env.action_space
print("else")
print(args.model)
self.obss_preprocessor = utils.ObssPreprocessor(
args.model, observation_space,
getattr(self.args, 'pretrained_model', None))
# Define actor-critic model
self.acmodel = utils.load_model(args.model, raise_not_found=False)
if self.acmodel is None:
if getattr(self.args, 'pretrained_model', None):
self.acmodel = utils.load_model(args.pretrained_model,
raise_not_found=True)
else:
self.acmodel = ACModel(self.obss_preprocessor.obs_space,
action_space, args.image_dim,
args.memory_dim, args.instr_dim,
not self.args.no_instr,
self.args.instr_arch,
not self.args.no_mem, self.args.arch)
self.obss_preprocessor.vocab.save()
utils.save_model(self.acmodel, args.model)
self.acmodel.train()
if torch.cuda.is_available():
self.acmodel.cuda()
self.optimizer = torch.optim.Adam(self.acmodel.parameters(),
self.args.lr,
eps=self.args.optim_eps)
self.scheduler = torch.optim.lr_scheduler.StepLR(self.optimizer,
step_size=100,
gamma=0.9)
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
def main(exp, argv):
os.environ["BABYAI_STORAGE"] = exp.results_directory()
# Parse arguments
parser = ArgumentParser()
parser.add_argument("--algo",
default='ppo',
help="algorithm to use (default: ppo)")
parser.add_argument("--discount",
type=float,
default=0.99,
help="discount factor (default: 0.99)")
parser.add_argument("--reward-scale",
type=float,
default=20.,
help="Reward scale multiplier")
parser.add_argument(
"--gae-lambda",
type=float,
default=0.99,
help="lambda coefficient in GAE formula (default: 0.99, 1 means no gae)"
)
parser.add_argument("--value-loss-coef",
type=float,
default=0.5,
help="value loss term coefficient (default: 0.5)")
parser.add_argument("--max-grad-norm",
type=float,
default=0.5,
help="maximum norm of gradient (default: 0.5)")
parser.add_argument("--clip-eps",
type=float,
default=0.2,
help="clipping epsilon for PPO (default: 0.2)")
parser.add_argument("--ppo-epochs",
type=int,
default=4,
help="number of epochs for PPO (default: 4)")
parser.add_argument(
"--save-interval",
type=int,
default=50,
help=
"number of updates between two saves (default: 50, 0 means no saving)")
parser.add_argument("--workers",
type=int,
default=8,
help="number of workers for PyTorch (default: 8)")
parser.add_argument("--max-count",
type=int,
default=1000,
help="maximum number of frames to run for")
parser.add_argument("--sample_duration",
type=float,
default=0.5,
help="sampling duration")
parser.add_argument("--cuda",
action="store_true",
default=False,
help="whether to use cuda")
args = parser.parse_args(argv)
utils.seed(args.seed)
torch_settings = init_torch(
seed=args.seed,
cuda=args.cuda,
workers=args.workers,
)
# Generate environments
envs = []
for i in range(args.procs):
env = gym.make(args.env)
env.seed(100 * args.seed + i)
envs.append(env)
# Define model name
suffix = datetime.datetime.now().strftime("%y-%m-%d-%H-%M-%S")
instr = args.instr_arch if args.instr_arch else "noinstr"
mem = "mem" if not args.no_mem else "nomem"
model_name_parts = {
'env': args.env,
'algo': args.algo,
'arch': args.arch,
'instr': instr,
'mem': mem,
'seed': args.seed,
'info': '',
'coef': '',
'suffix': suffix
}
default_model_name = "{env}_{algo}_{arch}_{instr}_{mem}_seed{seed}{info}{coef}_{suffix}".format(
**model_name_parts)
if args.pretrained_model:
default_model_name = args.pretrained_model + '_pretrained_' + default_model_name
args.model = args.model.format(
**model_name_parts) if args.model else default_model_name
utils.configure_logging(args.model)
logger = logging.getLogger(__name__)
# Define obss preprocessor
if 'emb' in args.arch:
obss_preprocessor = utils.IntObssPreprocessor(
args.model, envs[0].observation_space, args.pretrained_model)
else:
obss_preprocessor = utils.ObssPreprocessor(args.model,
envs[0].observation_space,
args.pretrained_model)
# Define actor-critic model
# acmodel = utils.load_model(args.model, raise_not_found=False)
acmodel = None
if acmodel is None:
if args.pretrained_model:
acmodel = utils.load_model(args.pretrained_model,
raise_not_found=True)
else:
acmodel = ACModel(obss_preprocessor.obs_space,
envs[0].action_space, args.image_dim,
args.memory_dim, args.instr_dim,
not args.no_instr, args.instr_arch,
not args.no_mem, args.arch)
obss_preprocessor.vocab.save()
# utils.save_model(acmodel, args.model)
if torch_settings.cuda:
acmodel.cuda()
# Define actor-critic algo
reshape_reward = lambda _0, _1, reward, _2: args.reward_scale * reward
if args.algo == "ppo":
algo = babyai.rl.PPOAlgo(
envs, acmodel, args.frames_per_proc, args.discount, args.lr,
args.beta1, args.beta2, args.gae_lambda, args.entropy_coef,
args.value_loss_coef, args.max_grad_norm, args.recurrence,
args.optim_eps, args.clip_eps, args.ppo_epochs, args.batch_size,
obss_preprocessor, reshape_reward)
else:
raise ValueError("Incorrect algorithm name: {}".format(args.algo))
# When using extra binary information, more tensors (model params) are initialized compared to when we don't use that.
# Thus, there starts to be a difference in the random state. If we want to avoid it, in order to make sure that
# the results of supervised-loss-coef=0. and extra-binary-info=0 match, we need to reseed here.
utils.seed(args.seed)
# Restore training status
status_path = os.path.join(utils.get_log_dir(args.model), 'status.json')
if os.path.exists(status_path):
with open(status_path, 'r') as src:
status = json.load(src)
else:
status = {'i': 0, 'num_episodes': 0, 'num_frames': 0}
# # Define logger and Tensorboard writer and CSV writer
# header = (["update", "episodes", "frames", "FPS", "duration"]
# + ["return_" + stat for stat in ['mean', 'std', 'min', 'max']]
# + ["success_rate"]
# + ["num_frames_" + stat for stat in ['mean', 'std', 'min', 'max']]
# + ["entropy", "value", "policy_loss", "value_loss", "loss", "grad_norm"])
# if args.tb:
# from tensorboardX import SummaryWriter
# writer = SummaryWriter(utils.get_log_dir(args.model))
# csv_path = os.path.join(utils.get_log_dir(args.model), 'log.csv')
# first_created = not os.path.exists(csv_path)
# # we don't buffer data going in the csv log, cause we assume
# # that one update will take much longer that one write to the log
# csv_writer = csv.writer(open(csv_path, 'a', 1))
# if first_created:
# csv_writer.writerow(header)
# Log code state, command, availability of CUDA and model
babyai_code = list(babyai.__path__)[0]
try:
last_commit = subprocess.check_output(
'cd {}; git log -n1'.format(babyai_code),
shell=True).decode('utf-8')
logger.info('LAST COMMIT INFO:')
logger.info(last_commit)
except subprocess.CalledProcessError:
logger.info('Could not figure out the last commit')
try:
diff = subprocess.check_output('cd {}; git diff'.format(babyai_code),
shell=True).decode('utf-8')
if diff:
logger.info('GIT DIFF:')
logger.info(diff)
except subprocess.CalledProcessError:
logger.info('Could not figure out the last commit')
logger.info('COMMAND LINE ARGS:')
logger.info(args)
logger.info("CUDA available: {}".format(torch.cuda.is_available()))
logger.info(acmodel)
# Train model
total_start_time = time.time()
best_success_rate = 0
best_mean_return = 0
test_env_name = args.env
wrapper = iteration_wrapper(
exp,
sync=torch_settings.sync,
max_count=args.max_count,
sample_duration=args.sample_duration,
)
# while status['num_frames'] < args.frames:
while True:
with wrapper() as it:
# Update parameters
if wrapper.done():
break
update_start_time = time.time()
logs = algo.update_parameters()
update_end_time = time.time()
it.set_count(logs["num_frames"])
it.log(loss=logs["loss"], )
env = RGBImgPartialObsWrapper(env)
env.seed(100 * args.seed + i)
envs.append(env)
# rolling mean success rate
ReturnsBuffer = deque(maxlen=5_000)
# Define obss preprocessor
if 'emb' in args.arch:
obss_preprocessor = utils.IntObssPreprocessor(args.model,
envs[0].observation_space,
args.pretrained_model)
else:
obss_preprocessor = utils.ObssPreprocessor(
instr_arch=args.instr_arch,
model_name=args.model,
obs_space=envs[0].observation_space,
load_vocab_from=args.pretrained_model)
# Define actor-critic model
acmodel = utils.load_model(args.model, raise_not_found=False)
if acmodel is None:
if args.pretrained_model:
acmodel = utils.load_model(args.pretrained_model, raise_not_found=True)
elif 'gie' in args.instr_arch:
acmodel = ACModel(obss_preprocessor.obs_space,
envs[0].action_space,
args.image_dim,
args.memory_dim,
args.instr_dim,
not args.no_instr,
mem = "mem" if not args.no_mem else "nomem"
fakerewardtxt = '' if args.fake_reward == 0. else 'fr{}_'.format(args.fake_reward)
default_model_name = "{}_{}_{}_{}_{}_seed{}_{}{}".format(args.env,
args.algo,
args.arch,
instr,
mem,
args.seed,
fakerewardtxt,
suffix)
model_name = args.model or default_model_name
# Define obss preprocessor
obss_preprocessor = utils.ObssPreprocessor(model_name, envs[0].observation_space)
# Define actor-critic model
acmodel = utils.load_model(model_name, raise_not_found=False)
if acmodel is None:
acmodel = ACModel(obss_preprocessor.obs_space, envs[0].action_space,
args.instr_model, not args.no_mem, args.arch)
if torch.cuda.is_available():
acmodel.cuda()
# Define actor-critic algo
if args.algo == "a2c":
algo = torch_rl.A2CAlgo(envs, acmodel, args.frames_per_proc, args.discount, args.lr, args.gae_tau,
args.entropy_coef, args.value_loss_coef, args.max_grad_norm, args.recurrence,
'student_obs_type': args.student_obs_type,
'dropout_rate': args.dropout,
'class_weights': args.class_weights
}
default_model_name = "{prefix}_{env}_{comm_freq}_{message_length}_{vocab_size}_{teacher_obs}_{student_obs_type}_{dropout_rate}_{class_weights}".format(
**model_name_parts)
if args.pretrained_model:
default_model_name = args.pretrained_model + '_pretrained_' + default_model_name
args.model = args.model.format(
**model_name_parts) if args.model else default_model_name
utils.configure_logging(args.model)
logger = logging.getLogger(__name__)
preprocess_obss_student = utils.ObssPreprocessor(args.model, "egocentric",
args.pretrained_model)
preprocess_obss_teacher = utils.ObssPreprocessor(args.model, args.teacher_obs,
args.pretrained_model,
envs[0].room_size)
#Define actor-critic model
acmodel = utils.load_model(args.model, raise_not_found=False)
if acmodel is None:
if args.pretrained_model:
acmodel = utils.load_model(args.pretrained_model, raise_not_found=True)
else:
teacher = Teacher(preprocess_obss_teacher.obs_space,
args.message_length, args.vocab_size, args.image_dim,
args.memory_dim, args.instr_dim,
args.comm_decoder_dim, not args.no_instr,
args.instr_arch, not args.no_mem, args.arch)
def __init__(self, args):
"""
:param args:
"""
super(MetaLearner, self).__init__()
self.update_lr = args.update_lr
self.meta_lr = args.meta_lr
self.task_num = args.task_num
self.args = args
utils.seed(self.args.seed)
self.env = gym.make(self.args.env)
demos_path = utils.get_demos_path(args.demos,
args.env,
args.demos_origin,
valid=False)
demos_path_valid = utils.get_demos_path(args.demos,
args.env,
args.demos_origin,
valid=True)
logger.info('loading demos')
self.train_demos = utils.load_demos(demos_path)
logger.info('loaded demos')
# if args.episodes:
# if args.episodes > len(self.train_demos):
# raise ValueError("there are only {} train demos".format(len(self.train_demos)))
# self.train_demos = self.train_demos[:args.episodes]
self.val_demos = utils.load_demos(demos_path_valid)
# if args.val_episodes > len(self.val_demos):
# logger.info('Using all the available {} demos to evaluate valid. accuracy'.format(len(self.val_demos)))
self.val_demos = self.val_demos[:self.args.val_episodes]
observation_space = self.env.observation_space
action_space = self.env.action_space
print(args.model)
self.obss_preprocessor = utils.ObssPreprocessor(
args.model, observation_space,
getattr(self.args, 'pretrained_model', None))
# Define actor-critic model
# self.net = utils.load_model(args.model, raise_not_found=False)
# if self.net is None:
# if getattr(self.args, 'pretrained_model', None):
# self.net = utils.load_model(args.pretrained_model, raise_not_found=True)
# else:
self.net = ACModel(self.obss_preprocessor.obs_space, action_space,
args.image_dim, args.memory_dim, args.instr_dim,
not self.args.no_instr, self.args.instr_arch,
not self.args.no_mem, self.args.arch)
self.obss_preprocessor.vocab.save()
# utils.save_model(self.net, args.model)
self.fast_net = copy.deepcopy(self.net)
self.net.train()
self.fast_net.train()
if torch.cuda.is_available():
self.net.cuda()
self.fast_net.cuda()
self.optimizer = torch.optim.SGD(self.fast_net.parameters(),
lr=self.args.update_lr)
# self.scheduler = torch.optim.lr_scheduler.StepLR(self.optimizer, step_size=100, gamma=0.9)
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.meta_optim = optim.Adam(self.net.parameters(), lr=self.meta_lr)
