Manytasks

A lightweight tool for deploying many tasks automatically, without any modification to your code.

• Manytasks
  • Installation
  • Quick Example
  • Sample Configuration
    • Syntax Sugar
  • Design Philosophy
  • Analysis
  • History

Installation

I recommend you install from github to get the newest features:

pip install git+https://github.com/dugu9sword/manytasks.git

You can also install the package (maybe outdated) from pypi:

pip install manytasks [not recommended]

Quick Example

cd examples/python

# configuration is stored in tasks.json
manytasks run tasks

All running logs are stored in tasks.logs.

• The running log of the manytasks is written into task.logs/status.txt

• The IO stream of tasks (e.g. print(), Exception, Error, ...) are redirected to task.logs/task-[index].txt.

Sample Configuration

For more complex cases, see ADVANCED_CASES.md.

{
  "executor": "python some.py",    
  "cuda": [4, 5, 6, 7],
  "concurrency": 4,           # num of tasks to run in parallel
  "cuda_per_task": 1,
  "configs": {
    # basic configurations
    "==base==": [          
      "arg0",
      "--a", [50, 100],       # `--a` takes value from [50, 100]
      "--name", "a_${a}"      # "${a}" refers to the value of `--a`
    ],
    # more disjoint configurations
    "==more==": [
      [ "--c1", [1, 2] ],            # [1, 2]
      [ "--c2", "$<x|y>.$<1:4>" ],   # [x.1, x.2, x.3, y.1, y.2, y.3]
    ]
  }
}

which yields:

---  ----  ---  ------  ----  ----
idx  __1   --a  --name  --c1  --c2
0    arg0  50   a_50    1     -
1    arg0  50   a_50    2     -
2    arg0  100  a_100   1     -
3    arg0  100  a_100   2     -
4    arg0  50   a_50    -     x.1
5    arg0  50   a_50    -     x.2
6    arg0  50   a_50    -     x.3
7    arg0  50   a_50    -     y.1
8    arg0  50   a_50    -     y.2
9    arg0  50   a_50    -     y.3
10   arg0  100  a_100   -     x.1
11   arg0  100  a_100   -     x.2
12   arg0  100  a_100   -     x.3
13   arg0  100  a_100   -     y.1
14   arg0  100  a_100   -     y.2
15   arg0  100  a_100   -     y.3
---  ----  ---  ------  ----  ----

Syntax Sugar

The syntax sugar makes the enumeration of arguments more easier.

Type	Example Input	Example Output
list	$<1:6>	[1, 2, 3, 4, 5]
list (with step)	$<1:6:2>	[1, 3, 5]
list (with zero-pad)	$<1:6:2;3>	[001, 003, 005]
files	$<files:/home/*.py>	[/home/foo.py, /home/bar.py, ...]
files (without path)	$<files:/home/*.py;nameonly>	[foo.py, bar.py, ...]
lines	$<lines:urls.txt>	[baidu.com, google.com, ...]
set	$<a|b|c>	[a, b, c]
composition	x_$<1:3;3>.txt	[x_001.txt, x_002.txt]
composition (more)	logs/$<a|b>.$<1:3>	[logs/a.1, logs/a.2, logs/b.1, logs/b.2]

Design Philosophy

Q: Why not use other open-source tools for hyper-parameter tuning, such as optuna, hydra, wandb?

A: The first time I developed this tool is 2019-1-1, when most of those tools (except optuna) were not born yet. What's more,

• These tools are heavy, which means:
  • You need to modify your code (add several lines of import xxx; xxx.foo(); xxx.bar();) to integrate their functionalities.
  • You may have to replace your print(...) or log(...) function with theirs, sometimes your results may be logged to their server.
  • These tools are highly binded with python, which means they may fail if your code is written in perl (such as mosetokenizer in machine translation), bash (such as your code for preprocessing), etc.
• manytasks is a lightweight tool which helps you deploy many tasks without any modification to your code, all logs will be saved in your machine.

Q:When should you use other open-source tools for hyper-parameter tuning, such as optuna, hydra, wandb?

A: For complex cases when you would like to enjoy their power of hyper-parameter tuning.

Analysis

See Analysis.md.

History

See History.md.