WIP: eval-update smoke test #1104
Conversation
host-requirements.txt
Outdated
| @@ -26,8 +26,15 @@ tensorflow-gpu==1.15.0 | |||
|
|
|||
| ### Begin Pytorch required libraries | |||
| pytorch==1.7.0 | |||
| torchvision==0.8.0 | |||
| torchvision==0.7.0 | |||
There was a problem hiding this comment.
happened to glance at this PR and was curious why the change to torchvision version? For reference, corresponding PyTorch/torchvision versions are listed in this table
There was a problem hiding this comment.
I was curious, too; thanks for the reference document. There was a pip version error when running --no-docker which may not be the proper consumer of host-requirements.txt. I am reverting that version regression and will better explore what the issue was.
There was a problem hiding this comment.
I believe the packages added in 4e21bcd are only needed for the ASR scenario. Perhaps we should add a comment making this clear
There was a problem hiding this comment.
Probably should, but let's keep that out of this PR.
|
@mwartell can you update for black and flake8? |
| logger.info(f"Applying internal {defense_type} defense to estimator") | ||
| estimator = config_loading.load_defense_internal(defense_config, estimator) | ||
|
|
||
| if model_config["fit"]: |
There was a problem hiding this comment.
for the sake of modularity, would it make sense to have a separate method encapsulating model training? _load_defense() isn't the first place I'd expect training to live
|
Since |
|
I would like to avoid those sort of append operations, which will quickly kill memory for large datasets. These are there mainly for easy interactive use for a single sample. |
|
Moved to #1114 |
Contains fixes for typo level errors in davidslater/eval-update (b9ae00c) found through scenario testing.
Incorporates davidslater/eval-update since that is the branch base.