Downloading an artifact deletes the summary of the last run in the sweep

[mahmudfami]

Hi everyone,
I’m encountering a problem with downloading model artifacts created in a wandb sweep.
We download the model artifact after the sweep using the api as usual:

api = wandb.Api()
artifact=api.artifact(model_name)
model_dir=artifact.download()

After the download the last run’s summary of the sweep, which created this artifact, gets deleted. The metrics associated with that run also disappear from the sweep view.

Has somebody a similar issue and can help?

[joana-marie]

Hi @mahmudfami ,
Is it possible if you can share your workspace link for us to review and what version of wandb SDK you are in now?

[joana-marie]

Hi @mahmudfami , since we have not heard back from you we are going to close this request. If you would like to re-open the conversation, please let us know!

[mahmudfami]

Hi @joana-marie , I’m sorry for the late reply.

Unfortunately, we cannot share our workspace.
We are using version 0.16.1. Maybe you can reproduce the error by running a sweep and afterwards downloading the best model artifact using the following code:

sweep_id = wandb.sweep(sweep_config, project="<project-name>")
wandb.agent(sweep_id, train)
  
# Get best model of sweep
api = wandb.Api()
sweep = api.sweep(f"<project-name>/{sweep_id}")
best_run = sweep.best_run()

artifacts = best_run.logged_artifacts()

model_artifact = None
for artifact in artifacts:
    if 'model' in artifact.type:  
        model_artifact = artifact
        break

if model_artifact != None:
  model_artifact_name = model_artifact.name
  print(f"Best Model: {model_artifact_name}")
else:
  warnings.warn(f"No model was found")
  
artifact=api.artifact(model_artifact_name)
model_dir=artifact.download()

[mahmudfami]

This is how the sweep graph looks afterwards:

[joana-marie]

Hi @mahmudfami ,
Sure thing, no worries about the workspace link. We are still doing some stuffs about this case. Could you try to downgrade to 0.15.12 (pip iinstall --upgrade wandb==0.15.12 ) and try again?

[mohammadbakir]

Hi @mahmudfami , we wanted to let you know we have a reproduction of this issue and are working on a fix with our next release. We will keep you updated.