Is it possible to have two runs concurrently in one script?

Description

I want to have two runs from two projects in one script. Is this supported, or do you have any viable alternatives for this?

This is different from this commonly asked question in that I also want another run outside the loop.

What I Did

meta_run = wandb.init(project="meta")

for ... # Meta loop
    individual_run= wandb.init(project="individual")
    for ... # Training loop
        individual_run.log(...)
    individual_run.join()
    meta_run.log()

[seungjaeryanlee]

@cvphelps, like this?

import random
import wandb

wandb.init(project="delete-this")

for n_units in [1, 2, 3]:
    for epoch_i in [1, 2, 3, 4, 5]:
        wandb.log({"train/loss": epoch_i * 100 - random.randint(-10, 10)})
        wandb.log({"valid/loss": epoch_i * 100 - random.randint(-20, 20)})
    
    wandb.log({
        "n_units": n_units,
        "test/loss": n_units * 10 - random.randint(-5, 5),
    })

This solution does not work for me since I can’t plot the individual runs separately. These are the plots I get:

Is there a way to plot each individual run separately?

By the way, thank you for your swift responses! You are awesome 😄

[vanpelt]

@seungjaeryanlee if your n_units actually correspond to different models you can start 3 different runs:

import random
import wandb

for n_units in [1, 2, 3]:
    wandb.init(project="delete-this", config={"n_units": n_units}, reinit=True)
    for epoch_i in [1, 2, 3, 4, 5]:
        wandb.log({"train/loss": epoch_i * 100 - random.randint(-10, 10)}, step=epoch_i)
        wandb.log({"valid/loss": epoch_i * 100 - random.randint(-20, 20)}, step=epoch_i)
    wandb.log({
        "test/loss": n_units * 10 - random.randint(-5, 5),
    })