Codecov Report

Merging #3465 (6be7a4e) into master (1bc04df) will increase coverage by 0.01%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #3465      +/-   ##
==========================================
+ Coverage   81.62%   81.63%   +0.01%     
==========================================
  Files         236      236              
  Lines       29113    29116       +3     
==========================================
+ Hits        23763    23770       +7     
+ Misses       5350     5346       -4     

Flags with carried forward coverage won't be shown. Click here to find out more.

BTW, tried this fix for wandb.agent that start a run a function that executes multiprocessing and it breaks:
(use case for this is if it is a multiprocessing training job with torch.multiprocessing.spawn for example)

import wandb
import multiprocessing as mp

def train_fn(rank):
    with wandb.init() as run:
        run.log({"rank": rank})

def agent_fn():
    procs = [mp.Process(target=train_fn, args=(rank,)) for rank in range(2)]
    for p in procs:
        p.start()

    for p in procs:
        p.join()

def sweep_fn():
    config = dict(
        method="random",
        parameters=dict(
            param0=dict(values=[2]),
            param1=dict(values=[0, 1, 4]),
            param2=dict(values=[0, 0.5, 1.5]),
            epochs=dict(value=4),
            )
        )
    sweep_id = wandb.sweep(config)
    wandb.agent(sweep_id, function=agent_fn, count=1)

if __name__ == "__main__":
    wandb.require("service")
    sweep_fn()

which is an issue that our growth asked for: WB-8808

(so from what I understand sweep job doesn't allow multiprocessing)

@kptkin
Great example.... that we should support... someday :)

right now the thread context called by wandb.agent() gets a hardcoded runid, so this example wont work with service or without.

We have a workaround at:
https://stackoverflow.com/questions/63469762/weightsbiases-sweep-keras-k-fold-validation

We will be able to support this eventually, but we need to figure out a better way to pass around the RUN_ID from the sweep or make it easier for users to spawn a new context inside a sweep that is independent of a sweep.