Model name convention

Format

<host>-<gpu config>-<core model name>-<internet source name>-<if modified use dax>-<timestamp>

Explanation of tags:

• host: where gives gpus
• gpu config: x gpu name y GiB
• core model name: name of the largest model or name of the most important model or name of ensemble system.
• internet source: name that can be searched easily on the internet to find core model name
• dax: to indicate that I (Vu Dinh Anh) have modified model or hyper-parameters. If there is no, it means that I used exact model or hyper-parameters.
• timestamp: year, month, day, hour, minute ~ YY_MM_DD_HH_mm

Rule: if there is a dash in a tag, replace it by underscore.

Example

ICT6-1_K80_12GiB-UNITER-vladsandulescu-dax-2021_04_30_09_39

This means that on ict 6, with 1 NVIDIA K80 12 GiB, UNITER model from vladsandulescu has been modified to train on 2021/04/30 at 09:39.

Parser

Split by dash to get the tag information.

Model reproduce instruction file

File name

Follow # Model name convention. File extension: .md.

File content

Detailed instruction to reproduce model from setup to train the model as well as expected outcome.

Example

Filename: ICT6-1_K80_12GiB-UNITER-vladsandulescu-dax-2021_04_30_09_39.md

Content:

# I install like this

# Then I run like this

# Then I get these results

# Then I conclude like this

utilities.py

This file provides 2 functions.

def parse_model_name(model_name, verbose=True):
    """To parse model name into json/dict

    Parameters:
        model_name: a string
        verbose: a bool

    Returns:
        a json/dict
    """

def pprint_md(markdown_file):
    """To pretty print a markdown file with pygmentize

    Run `pip install Pygments` if there is none

    Parameters:
        markdown_file: a path to markdown file

    Returns: 
        none
    """

Run utilities.py to see the demo.