def setup_logger(config):
log_path = config.get("log_path", os.getcwd())
if is_main_process():
backends = [
TensorBoardBackend(verbosity=dllogger.Verbosity.VERBOSE,
log_dir=log_path),
JSONStreamBackend(verbosity=dllogger.Verbosity.VERBOSE,
filename=os.path.join(log_path, "log.json")),
AggregatorBackend(verbosity=dllogger.Verbosity.VERBOSE,
agg_dict={"loss": AverageMeter}),
StdOutBackend(
verbosity=dllogger.Verbosity.DEFAULT,
step_format=empty_step_format,
metric_format=no_string_metric_format,
prefix_format=empty_prefix_format,
),
]
logger = Logger(backends=backends)
else:
logger = Logger(backends=[])
container_setup_info = get_framework_env_vars()
logger.log(step="PARAMETER",
data=container_setup_info,
verbosity=dllogger.Verbosity.DEFAULT)
logger.metadata("loss", {
"unit": "nat",
"GOAL": "MINIMIZE",
"STAGE": "TRAIN"
})
logger.metadata("val_loss", {
"unit": "nat",
"GOAL": "MINIMIZE",
"STAGE": "VAL"
})
return logger
class dllogger_class():
def format_step(self, step):
if isinstance(step, str):
return step
elif isinstance(step, int):
return "Iteration: {} ".format(step)
elif len(step) > 0:
return "Iteration: {} ".format(step[0])
else:
return ""
def __init__(self, log_path="bert_dllog.json"):
self.logger = Logger([
StdOutBackend(Verbosity.DEFAULT, step_format=self.format_step),
JSONStreamBackend(Verbosity.VERBOSE, log_path),
])
self.logger.metadata("mlm_loss", {
"format": ":.4f",
"GOAL": "MINIMIZE",
"STAGE": "TRAIN"
})
self.logger.metadata("nsp_loss", {
"format": ":.4f",
"GOAL": "MINIMIZE",
"STAGE": "TRAIN"
})
self.logger.metadata("avg_loss_step", {
"format": ":.4f",
"GOAL": "MINIMIZE",
"STAGE": "TRAIN"
})
self.logger.metadata("total_loss", {
"format": ":.4f",
"GOAL": "MINIMIZE",
"STAGE": "TRAIN"
})
self.logger.metadata("loss", {
"format": ":.4f",
"GOAL": "MINIMIZE",
"STAGE": "TRAIN"
})
self.logger.metadata("f1", {
"format": ":.4f",
"GOAL": "MINIMIZE",
"STAGE": "VAL"
})
self.logger.metadata("precision", {
"format": ":.4f",
"GOAL": "MINIMIZE",
"STAGE": "VAL"
})
self.logger.metadata("recall", {
"format": ":.4f",
"GOAL": "MINIMIZE",
"STAGE": "VAL"
})
self.logger.metadata("mcc", {
"format": ":.4f",
"GOAL": "MINIMIZE",
"STAGE": "VAL"
})
self.logger.metadata("exact_match", {
"format": ":.4f",
"GOAL": "MINIMIZE",
"STAGE": "VAL"
})
self.logger.metadata(
"throughput_train",
{
"unit": "seq/s",
"format": ":.3f",
"GOAL": "MAXIMIZE",
"STAGE": "TRAIN"
},
)
self.logger.metadata(
"throughput_inf",
{
"unit": "seq/s",
"format": ":.3f",
"GOAL": "MAXIMIZE",
"STAGE": "VAL"
},
)
l = Logger(
[
StdOutBackend(Verbosity.DEFAULT, step_format=format_step),
JSONStreamBackend(Verbosity.VERBOSE, "tmp.json"),
]
)
# You can log metrics in separate calls
l.log(step="PARAMETER", data={"HP1": 17}, verbosity=Verbosity.DEFAULT)
l.log(step="PARAMETER", data={"HP2": 23}, verbosity=Verbosity.DEFAULT)
# or together
l.log(step="PARAMETER", data={"HP3": 1, "HP4": 2}, verbosity=Verbosity.DEFAULT)
l.metadata("loss", {"unit": "nat", "GOAL": "MINIMIZE", "STAGE": "TRAIN"})
l.metadata("val.loss", {"unit": "nat", "GOAL": "MINIMIZE", "STAGE": "VAL"})
l.metadata(
"speed",
{"unit": "speeds/s", "format": ":.3f", "GOAL": "MAXIMIZE", "STAGE": "TRAIN"},
)
for epoch in range(0, 2):
for it in range(0, 10):
l.log(
step=(epoch, it),
data={"loss": 130 / (1 + epoch * 10 + it)},
verbosity=Verbosity.DEFAULT,
)
if it % 3 == 0: