Move no_grad() to WorkerTemplate._update_all_metrics() - Githubissues

amjltc295 / pytorch-golden-template

PyTorch Golden Template (under development)

13 stars 4 forks source link

Move no_grad() to WorkerTemplate._update_all_metrics() #64

Closed amjltc295 closed 5 years ago

amjltc295 commented 5 years ago

Why do we need this PR?

If no_grad() is not added in the _update() function of a metric, there would be GPU memory leaking issue.

How is it implemented?

As no_grad() is required for all metrics, it is now moved to WorkerTemplate._update_all_metrics().

Other changes

Add comments for BaseMetric

PR readiness checklist

[x] Did it pass the Flake8 check?

[x] Did you test this PR?