Sweep configurations

Hi I have two questions regarding sweeps:

1. Is it possible to re-configure a sweep?

Consider the following simple example:

• I have a single hyperparameter - learning_rate.
• I create the following sweep configuration YAML file (named wandb_conf.yaml):

program: main.py
method: grid
metric:
  name: val_acc
  goal: maximize
parameters:
  learning_rate:
    distribution: categorical
    values: [1e-5, 1e-4, 1e-3, 1e-2]

• I create a new sweep using the following command: wandb sweep wandb_conf.yaml
• New sweep is created with ID SWEEP_EXAMPLE_ID
• I start the sweep by running wandb agent SWEEP_EXAMPLE_ID
• After the agent runs several experiments, it’s obvious that setting learning_rate to 1e-2 is bad for the model

What I would like to do is:

• Pause the sweep.
• Update the sweep’s configuration by removing the value 1e-2 from the list of values of the learning_rate parameter:

parameters:
  learning_rate:
    distribution: categorical
    values: [1e-5, 1e-4, 1e-3]

• Continue sweep

2. Is it possible to assign per-agent command line parameters? Let’s say that my program accepts the command line argument --gpus which sets the number of GPUs it uses. If I set this argument in the sweep’s YAML file, e.g.:

command:
      - "--gpus"
      - "4"

then, each agent that starts, will start an experiment using 4 GPUs.

This behavior is too strict. Let’s say I have two servers. Server A with 4 GPUs and server B with 8 GPUs. What I would like to be able to do is:

• On server A: `wandb agent SWEEP_ID --gpus 4
• On server B: `wandb agent SWEEP_ID --gpus 8

[issue-label-bot[bot]]

Issue-Label Bot is automatically applying the label question to this issue, with a confidence of 0.61. Please mark this comment with 👍 or 👎 to give our bot feedback!

Links: app homepage, dashboard and code for this bot.

[sydholl]

In the past year we’ve majorly reworked the CLI and UI for Weights & Biases. We’re closing issues older than 6 months. Please comment to reopen.