How to get wandb to pass arguments by position?-CodePudding

I am trying to explore the results of different parameter settings on my python script "train.py". For that, I use a wandb sweep. Each wandb agent executes the file "train.py" and passes some parameters to it. As per the wandb documentation (https://docs.wandb.ai/guides/sweeps/configuration#command), in case of e.g. two parameters "param1" and "param2" each agents starts the file with the command

/usr/bin/env python train.py --param1=value1 --param2=value2

However, "train.py" expects

/usr/bin/env python train.py value1 value2

and parses the parameter values by position. I did not write train.py and would like to not change it if possible. How can I get wandb to pass the values without "--param1=" in front?

CodePudding user response：

Don't think you can get positional arguments from W&B Sweeps. However, there's a little work around you can try that won't require you touching the train.py file.

You can create an invoker file, let's call it invoke.py. Now, you can use it get rid of the keyword argument names. Something like this might work:

import sys
import subprocess

if len(sys.argv[0]) <= 1:
  print(f"{sys.argv[0]} program_name param0=<param0> param1=<param1> ...")
  sys.exit(0)

program = sys.argv[1]
params = sys.argv[2:]

posparam = []
for param in params:
  _, val = param.split("=")
  posparam.append(val)

command = [sys.executable, program, *posparam]
process = subprocess.Popen(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
out, err = process.communicate()
sys.stdout.write(out.decode())
sys.stdout.flush()
sys.stderr.write(err.decode())
sys.stderr.flush()
sys.exit(process.returncode)

This allows you to invoke your train.py file as follows:

$ python3 invoke.py /path/to/train.py param0=0.001 param1=20 ...

Now to perform W&B sweeps you can create a command: section (reference) in your sweeps.yaml file while sweeping over the parameters param0 and param1. For example:

program: invoke.py
...
parameters:
  param0:
    distribution: uniform
    min: 0
    max: 1
  param1:
    distribution: categorical
    values: [10, 20, 30]
command:
 - ${env}
 - ${program}
 - /path/to/train.py
 - ${args_no_hyphens}