Skip to content

Latest commit

 

History

History
74 lines (63 loc) · 2.83 KB

log_format.md

File metadata and controls

74 lines (63 loc) · 2.83 KB
Raw

Lifelong Learning Logger Output Format

This document describes the l2logger output format in terms of the directory structure created and the contents of the files therein. This is version 1.1 of the log_output_format.

Directory Structure

The logger produces output according to the structure below. At instantiation, the logger takes in logging_base_dir to start as the top-level logging directory, creating it if it doesn't exist already.

logging_base_dir
└───scenario_dir
    │   logger_info.json
    │   scenario_info.json
    │
    └───worker-0
    │   └───block-0
    │   │    │   data-log.tsv
    │   ...
    │   │
    │   └───block-n
    │        │   data-log.tsv
    │
    ...
    └───worker-m
    │   └───block-0
    │   │    │   data-log.tsv
    │   ...
    │   │
    │   └───block-n
    │        │   data-log.tsv
    │   ...

From the directory structure, scenario_dir should be the name of a specific run of a given scenario (e.g. the scenario name followed by a timestamp, like simple-1600697775-609517). Within this folder are scenario_info.json and logger_info.json; see the 'Initialization' section in interface.md for explanation of their contents.

Still within this top-level scenario directory, each worker (e.g. thread or process) then gets its own folder to write logs to. Within a worker's folder, there is a folder for each block in the syllabus that the worker consumed experiences on (e.g. '0-train', '1-test', etc). Then, within each of these folders, is the data-log.tsv file containing the actual logged records for that (worker, block) pair.

The interface for providing this information to the logger, as well as details on the contents of the TSV files, is explained in more detail in interface.md as well as below.

data-log.tsv format

data-log.tsv is a tab-separated value file. It expects the first line to contain the column headers, then each subsequent line to contain values for each column separated by tabs. For a raw example of what this file should look like, see here.

Important to note about this format:

  • task_params is the result of calling json.dumps(...) on the params field in the record; as a result, the escaping of quotes and other special characters therein are rather particular.
  • timestamp is the result of formatting datetime.now() with the time format string %Y%m%dT%H%M%S.%f
    • This is effectively year, month, day followed by hour, minute, second with the fractional portion of the second as a single-precision float (e.g., 20201020T230415.363982).
  • All other fields are just dumped in as passed in, integers and strings alike, not needing any quotes or escape sequences