def __init__(self,
config,
experiments_prefix,
forced_debug=False,
logfile_name="log"):
"""Create a new Experiment instance.
Modified based on: https://github.com/ex4sperans/mag
Args:
logfile_name: str, naming for log file. This can be useful to
separate logs for different runs on the same experiment
experiments_prefix: str, a prefix to the path where
experiment will be saved
"""
# get all defaults
all_defaults = {}
for key in vars(config):
all_defaults[key] = get_base_parser().get_default(key)
self.default_config = all_defaults
config.resume = False
if not config.debug and not forced_debug:
if os.path.isdir(self.experiment_dir):
print("log exists: {}".format(self.experiment_dir))
config.resume = True
print(config)
self._makedir()
def experiment_dir(self):
if self.config.debug or self.forced_debug:
return "./"
else:
# get namespace for each group of args
arg_g = dict()
for group in get_base_parser()._action_groups:
group_d = {
a.dest: self.default_config.get(a.dest, None)
for a in group._group_actions
}
arg_g[group.title] = argparse.Namespace(**group_d)
# skip default value
identifier = ""
for key, value in sorted(vars(arg_g["model_configs"]).items()):
if getattr(self.config, key) != value:
identifier += key + str(getattr(self.config, key))
return os.path.join(self.experiments_prefix, identifier)
def __init__(self, config, experiments_prefix, logfile_name="log"):
"""Create a new Experiment instance.
Modified based on: https://github.com/ex4sperans/mag
Args:
logfile_name: str, naming for log file. This can be useful to
separate logs for different runs on the same experiment
experiments_prefix: str, a prefix to the path where
experiment will be saved
"""
# get all defaults
all_defaults = {}
for key in vars(config):
all_defaults[key] = get_base_parser().get_default(key)
self.default_config = all_defaults
# activation function
activate_list = ["Tanh", "ReLU", "Softplus", "LogSigmoid"]
map_activate = {act.lower(): act for act in activate_list}
if self.config.activate.lower() in map_activate:
self.config.activate = map_activate[self.config.activate.lower()]
else:
raise ValueError("Choose activation function among 'Tanh,'"
"ReLU, Softplus, logsigmoid'")
config.resume = False
if not config.debug:
if os.path.isdir(self.experiment_dir):
print("log exists: {}".format(self.experiment_dir))
config.resume = True
print(config)
self._makedir()
gen_init=gen_init,
theta_y=theta_y,
disc_init=netD_state,
experiment=e)
e.log.info(model)
model.train()
model.get_theta()
model.save()
pickle.dump([
theta_y, model.all_theta, model.all_delta, model.all_weights,
model.all_grad_norm, model.all_loss, model.all_tolerance
], open(e.experiment_dir + "/bayes_all+thetas+weights.pkl", "wb+"))
if __name__ == '__main__':
args = config.get_base_parser().parse_args()
with train_helper.experiment(args, args.save_prefix) as e:
np.random.seed(e.config.random_seed)
torch.manual_seed(e.config.random_seed)
e.log.info("*" * 25 + " ARGS " + "*" * 25)
e.log.info(args)
e.log.info("*" * 25 + " ARGS " + "*" * 25)
run(e)
iteration=true_it,
epoch=epoch + 1,
name="latest")
time_per_epoch = (e.elapsed_time / (epoch - start_epoch + 1))
time_in_need = time_per_epoch * (e.config.n_epoch - epoch - 1)
e.log.info("elapsed time: {:.2f}(h), "
"time per epoch: {:.2f}(h), "
"time needed to finish: {:.2f}(h)".format(
e.elapsed_time, time_per_epoch, time_in_need))
train_stats.reset()
if __name__ == '__main__':
PARSED_CONFIG = config.get_base_parser().parse_args()
def exit_handler(*args):
print(PARSED_CONFIG)
print("best dev bleu: {:.4f}, test bleu: {:.4f}".format(
BEST_DEV_BLEU, TEST_BLEU))
sys.exit()
train_helper.register_exit_handler(exit_handler)
with train_helper.Experiment(PARSED_CONFIG,
PARSED_CONFIG.save_prefix) as exp:
exp.log.info("*" * 25 + " ARGS " + "*" * 25)
exp.log.info(PARSED_CONFIG)
exp.log.info("*" * 25 + " ARGS " + "*" * 25)
