Getting started

This is a quick start guide to using the dbxlogger and dbx command line tool.

Install the logger

Install using pip directly from source:

pip install git+git://github.com/vladvelici/dbxlogger.git

Create an experiment

import dbxlogger
from dbxlogger import Exp

exp = Exp(kind="hello", repo=dbxlogger.get_repo("./output"))

# save the experiment to the repo
exp.save()

Create a logger

logger = exp.logger()

# log something
logger.log("event_name", {"some": "data"})

# or use the shortcut
logger("event_name", {"some": "data"})

Event names and the log context

Log events can be logically grouped by name. For instance, one can use a naming schema like epoch/<epoch_num>/start, epoch/<epoch_num>/end, epoch/<epoch_num>/train, epoch/<epoch_num>/validation and so on.

This enables us to easily see which events belong to which epoch and query the log accordingly.

Using a naming scheme like this is common. In fact, this is what dbxlogger uses for the torchbearer integration.

Implementing this can be done in a few ways:

# logger is the logger object
for epoch in logger.iter_at(range(max_epochs), lambda epoch: "epoch/%d" % epoch):
    logger.log("start", {})

# logs:
# {"event": "epoch/0/start"}
# {"event": "epoch/1/start"}
# {"event": "epoch/2/start"}
# ...
# {"event": "epoch/[max_epochs-1]/start"}

Or you can do this in a more manual way:

for epoch in range(max_epochs):
    logger.ctx.path = "epoch/%d" % epoch
    logger.log("start", {}) # logs the same thing as before