Initial version of classification reference scripts #819
Conversation
Also log the learning rate
Identified a bug in the reporting of the results. They need to be reduced between all processes
soumith
left a comment
soumith
left a comment
There was a problem hiding this comment.
reviewed the train script. didn't review classification/utills.py
|
|
||
|
|
||
| def setup_for_distributed(is_master): | ||
| """ |
There was a problem hiding this comment.
printing override seems fine, but torch.save override seems pretty sketchy. Maybe consider having a utils.save_on_master that you use, rather than monkey-patching torch.save
There was a problem hiding this comment.
sounds good.
This is a fundamental feature for distributed training, so it's better to have it right.
Codecov Report
@@ Coverage Diff @@
## master #819 +/- ##
=======================================
Coverage 51.58% 51.58%
=======================================
Files 34 34
Lines 3342 3342
Branches 536 536
=======================================
Hits 1724 1724
Misses 1486 1486
Partials 132 132Continue to review full report at Codecov.
|
3 participants
This PR introduces the foundations for reference training/evaluation scripts for torchvision.
The idea is that all pre-trained models will have corresponding training scripts / command-line arguments, so that reproducing a trained model should be straightforward.
This is not at its final version. I'll be merging this soon, and after adding segmentation and detection training/evaluation scripts, a lot of it will be refactored and included inside torchvision.