Python logging package for easy reproducible experimenting in research

smilelogging

Python logging package for easy reproducible experimenting in research.

Why you may need this package

This project is meant to provide an easy-to-use (as easy as possible) package to enable reproducible experimenting in research. Here is a struggling situation you may also encountered:

I am doing some project. I got a fatanstic idea some time (one week, one month, or even one year) ago. Now I am looking at the results of that experiment, but I just cannot reproduce them anymore. I cannot remember which script and what hyper-prarameters I used. Even worse, since then I've modified the code (a lot). I don't know where I messed it up...

If you do not use this package, usually, what you can do may be:

  • First, use Github to manage your code. Always run experiments after git commit.
  • Second, before each experiment, set up a unique experiment folder (with a unique ID to label that experiment -- we call it ExpID).
  • Third, when running an experiment, print your git commit ID (we call it CodeID) and arguments in the log.

Every result is uniquely binded with an ExpID, corresponding to a unique experiment folder. In that folder, CodeID and arguments are saved. So ideally, as long as we know the ExpID, we should be able to rerun the experiment under the same condition.

These steps are pretty simple, but if you implement them over and over again in each project, it can still be quite annoying. This package is meant to save you with basically 3~4 lines of code change.

Usage

Step 0: Install the package (>= python3.4)

# --upgrade to make sure you install the latest version
pip install smilelogging --upgrade

Step 1: Modify your code

Here we use the official PyTorch ImageNet example to give an example.

# 1. add this at the head of code
from smilelogging import Logger 

# 2. replace argument parser
parser = argparse.ArgumentParser(description='PyTorch ImageNet Training')  
==> change the above line to the following:
from smilelogging import argparser as parser

# 3. add logger and change print if necessary
args = parser.parse_args()
==> change the above line to the following 3:
args = parser.parse_args()
logger = Logger(args)
global print; print = logger.log_printer.logprint # change print function so that logs can be printed to a txt file

TIPS: overwriting the default python print func may not be a good practice, a better way may be logprint = logger.log_printer.logprint, and use it like logprint('Test accuracy: %.4f' % test_acc). This will print the log to a txt file at path log/log.txt.

Step 2: Run experiments

The original ImageNet training snippet is:

CUDA_VISIBLE_DEVICES=0 python main.py -a resnet18 [imagenet-folder with train and val folders]

Now, try this:

CUDA_VISIBLE_DEVICES=0 python main.py -a resnet18 [imagenet-folder with train and val folders] --project_name Scratch__resnet18__imagenet --screen_print

This snippet will set up an experiment folder under path Experiments/Scratch__resnet18__imagenet_XXX. That XXX thing is an ExpID automatically assigned by the time running this snippet. Below is an example on my PC:

Experiments/
└── Scratch__resnet18__imagenet_SERVER138-20211021-145936
    ├── gen_img
    ├── log
    │   ├── git_status.txt
    │   ├── gpu_info.txt
    │   ├── log.txt
    │   ├── params.yaml
    │   └── plot
    └── weights

CongratsYou're all set

As seen, there will be 3 folders automatically created: gen_img, weights, log. Log text will be saved in log/log.txt, arguments saved in log/params.yaml and in the head of log/log.txt. Below is an example of the first few lines of log/log.txt:

cd /home/wanghuan/Projects/TestProject
CUDA_VISIBLE_DEVICES=1 python main.py -a resnet18 /home/wanghuan/Dataset/ILSVRC/Data/CLS-LOC/ --project Scracth_resnet18_imagenet --screen_print

('arch': resnet18) ('batch_size': 256) ('cache_ignore': ) ('CodeID': f30e6078) ('data': /home/wanghuan/Dataset/ILSVRC/Data/CLS-LOC/) ('debug': False) ('dist_backend': nccl) ('dist_url': tcp://224.66.41.62:23456) ('epochs': 90) ('evaluate': False) ('gpu': None) ('lr': 0.1) ('momentum': 0.9) ('multiprocessing_distributed': False) ('note': ) ('pretrained': False) ('print_freq': 10) ('project_name': Scracth_resnet18_imagenet) ('rank': -1) ('resume': ) ('screen_print': True) ('seed': None) ('start_epoch': 0) ('weight_decay': 0.0001) ('workers': 4) ('world_size': -1)

[180853 22509 2021/10/21-18:08:54] ==> Caching various config files to 'Experiments/Scracth_resnet18_imagenet_SERVER138-20211021-180853/.caches'

Note, it tells us

  • (1) where is the code
  • (2) what snippet is used when running this experiment
  • (3) what arguments are used
  • (4) what is the CodeID -- useful when rolling back to prior code versions (git reset --hard )
  • (5) where the code files (*.py, *.json, *.yaml etc) are backuped -- note the log line "==> Caching various config files to ...". Ideally, CodeID is already enough to get previous code. Caching code files is a double insurance
  • (6) At the begining of each log line, the prefix "[180853 22509 2021/10/21-18:08:54]" is automatically added if the logprint func is used for print, where 180853 is short for the full ExpID SERVER138-20211021-180853, 22509 is the program pid (useful if you want to kill the job, e.g., kill -9 22509)

More explanantions about the folder setting

The weights folder is supposed to store the checkpoints during training; and gen_img is supposed to store the generated images during training (like in a generative model project). To use them in the code:

weights_path = logger.weights_path
gen_img_path = logger.gen_img_path

More explanantions about the arguments and more tips

  • --screen_print means the logs will also be print to the console (namely, your screen). If it is not used, the log will only be saved to log/log.txt, not printed to screen.
  • If you are debugging code, you may not want to create an experiment folder under Experiments. Then use --debug, for example:
CUDA_VISIBLE_DEVICES=0 python main.py -a resnet18 [imagenet-folder with train and val folders] --debug

This will save all the logs in Debug_Dir, instead of Experiments (Experiments is expected to store the formal experiment results).

TODO

  • Add training and testing metric (like accuracy, PSNR) plots.

Collaboration / Suggestions

Currently, this is still a baby project. Any collaboration or suggestions are welcome to Huan Wang (Email: [email protected]).

Owner
Huan Wang
B.E. and M.S. graduate from Zhejiang University, China. Now Ph.D. candidate at Northeastern, USA. I work on interpretable model compression.
Huan Wang
Similar Resources

Simple and versatile logging library for python 3.6 above

Simple and versatile logging library for python 3.6 above

Apr 5, 2022

Stand-alone parser for User Access Logging from Server 2012 and newer systems

Stand-alone parser for User Access Logging from Server 2012 and newer systems

KStrike Stand-alone parser for User Access Logging from Server 2012 and newer systems BriMor Labs KStrike This script will parse data from the User Ac

Sep 12, 2022

Logging system for the TPC software.

tpc_logger Logging system for the TPC software. The TPC Logger class provides a singleton for logging information within C++ code or in the python API

Jan 10, 2022

Outlog it's a library to make logging a simple task

outlog Outlog it's a library to make logging a simple task!. I'm a lazy python user, the times that i do logging on my apps it's hard to do, a lot of

Mar 5, 2022

metovlogs is a very simple logging library

metovlogs is a very simple logging library. Setup is one line, then you can use it as a drop-in print replacement. Sane and useful log format out of the box. Best for small or early projects.

Mar 1, 2022

The easy way to send notifications

The easy way to send notifications

See changelog for recent changes Got an app or service and you want to enable your users to use notifications with their provider of choice? Working o

Sep 25, 2022

A simple package that allows you to save inputs & outputs as .log files

wolf_dot_log A simple package that allows you to save inputs & outputs as .log files pip install wolf_dot_log pip3 install wolf_dot_log |Instructions|

Nov 16, 2021

Pretty-print tabular data in Python, a library and a command-line utility. Repository migrated from bitbucket.org/astanin/python-tabulate.

python-tabulate Pretty-print tabular data in Python, a library and a command-line utility. The main use cases of the library are: printing small table

Sep 20, 2022

Progressbar 2 - A progress bar for Python 2 and Python 3 - "pip install progressbar2"

Text progress bar library for Python. Travis status: Coverage: Install The package can be installed through pip (this is the recommended method): pip

Sep 8, 2022
Related tags
Python logging made (stupidly) simple
Python logging made (stupidly) simple

Loguru is a library which aims to bring enjoyable logging in Python. Did you ever feel lazy about configuring a logger and used print() instead?... I

Sep 27, 2022
Structured Logging for Python
Structured Logging for Python

structlog makes logging in Python faster, less painful, and more powerful by adding structure to your log entries. It's up to you whether you want str

Sep 19, 2022
A colored formatter for the python logging module
A colored formatter for the python logging module

Log formatting with colors! colorlog.ColoredFormatter is a formatter for use with Python's logging module that outputs records using terminal colors.

Sep 23, 2022
Colored terminal output for Python's logging module
Colored terminal output for Python's logging module

coloredlogs: Colored terminal output for Python's logging module The coloredlogs package enables colored terminal output for Python's logging module.

Sep 21, 2022
A cool logging replacement for Python.
A cool logging replacement for Python.

Welcome to Logbook Travis AppVeyor Supported Versions Latest Version Test Coverage Logbook is a nice logging replacement. It should be easy to setup,

Aug 27, 2022
Robust and effective logging for Python 2 and 3.
Robust and effective logging for Python 2 and 3.

Robust and effective logging for Python 2 and 3.

Sep 18, 2022
A basic logging library for Python.
A basic logging library for Python.

log.py ?? About: A basic logging library for Python with the capability to: save to files. have custom formats. have custom levels. be used instantiat

Jan 19, 2022
Small toolkit for python multiprocessing logging to file

Small Toolkit for Python Multiprocessing Logging This is a small toolkit for solving unsafe python mutliprocess logging (file logging and rotation) In

Nov 10, 2021
Beautifully colored, quick and simple Python logging
Beautifully colored, quick and simple Python logging

Python Quick Logging | QLogging Beautifully colored, quick and simple Python logging. This logger is based on Python logging package Screenshots: Term

Jun 14, 2022
A lightweight logging library for python applications

cakelog a lightweight logging library for python applications This is a very small logging library to make logging in python easy and simple. config o

Jan 5, 2022