Log wandb metrics when run is finished.

wandb_log_metrics_after_finished.py

"""
Load in a wandb run and updates its run history with a new metric.

DOCS: https://docs.wandb.ai/guides/track/public-api-guide#update-config-for-an-existing-run
"""
import wandb

api = wandb.Api()

entity, project = "kenevoldsen", "mnist-test"  # set to your entity and project
run_id = "q564eoml"  # set to your run id

# to get all runs use api.runs(enti
# ty + "/" + project)
run = api.run(entity + "/" + project + "/" + run_id)

type(run.history())
# pandas.core.frame.DataFrame
print(run.history())
#      _step      loss  _runtime    _timestamp  Accuracy
# 0        0  2.358063  0.870733  1.666799e+09       NaN
# 1        1  2.409657  0.913244  1.666799e+09       NaN
# 2        4  2.331224  0.925087  1.666799e+09       NaN

run.name # fetch the run name

# update run summary (not really want we want as we want to be able to plot the metrics over steps)
# however that is not possible with the public API yet (see: https://github.com/wandb/wandb/issues/2723)
run.summary.update({"Accuracy": 0.9})

run = wandb.init(project=project, id=run_id, resume=True)

for i in range(10):
    run.log({"epoch": i, "my_new_metric": 1.0})
    # note: you cannot use step here instead of epoch

run.finish()

# so for dfm it would be:
# run = wandb.init(project=project, id=run_id, resume=True)
# run.log({"train/global_step": step_for_scores, "scaneval_metric_1": score, ...})