a light weight experiment reproducibility toolset
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
example Add files via upload Dec 1, 2018
lite_tracer Add files via upload Dec 1, 2018
logos
LICENSE
LICENSE.GPL
README.md
setup.py

README.md

lite_tracer Logo


LiteTracer: a light weight experiment reproducibility toolset

LiteTracer acts as a drop-in replacement for argparse, and it can generate unique identifiers for experiments in addition to what argparse already does. Along with a reverse lookup tool, LiteTracer can trace-back the state of a project that generated any result tagged by the identifier. The identifiers are unique based on the combination of four factors:

  1. code version;
  2. un-committed code changes
  3. untracked files in the project;
  4. any command line arguments supplied at runtime.

As the name suggests, LiteTracer is designed to be as lightweight as possible. It is a minimalistic toolset and convention to enable reproducible experimental research, rather than a framework that one has to learn about.

To track:

  1. Instead of using argparse from argparse import ArgumentParser, use LTParser, e.g.:
from lite_tracer import LTParser
parser = LTParser("...")
parser.add_argument(...)
args = parser.parse_args()
  1. Then in any result file you save (tensorboard results included), include as part of filename: args.hash_code, for example:
result_path = './results/{}/{}'.format(args.data_name, args.hash_code)

NEVER manually change output filenames (e.g. use generated filenames directly in your latex source code)

Given hash code, to trace back to the exact configuration that produced a result:

By default, LTParser saves tracking information to ./lt_records/<args.hash_code>, which has three things:

settings_<args.hash_code>.txt which has all arguments used for the experiments (command line supplied merged with defaults), as well as some dynamically collected information such as git version information

diff.patch any source code change from last committed version

untracked/ any untracked and not ingored files/folders in the project dir

To find all results with certain param settings:

lite_trace.py --include [[[PARAM1:VAL1] PARAM2:VAL2] ...] --exclude [[[PARAM1:VAL1] PARAM2:VAL2] ...]

See a complete example in example/lite_tracer_example.py

Example search: lite_trace.py --exclude bsz:12 git_label:f6afeb8 --include sgd