add util for loss spike save and decode. #1044
Conversation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #1044 +/- ##
==========================================
- Coverage 78.53% 78.25% -0.29%
==========================================
Files 187 191 +4
Lines 17336 17784 +448
==========================================
+ Hits 13615 13916 +301
- Misses 3721 3868 +147 ☔ View full report in Codecov by Sentry. |
|
You need to format your commits to pass the test of atorch-pre-commit. |
There was a problem hiding this comment.
Use English for document writing and with readable format.
|
|
||
|
|
||
| class TokenLossSpike(LossSpikeBase): | ||
| def save_loss(self, file_name, cur_loss, cur_iter, losses_str, sample_infos_str): |
There was a problem hiding this comment.
What is the relationship between cur_loss and losses_str, or cur_iter and sample_infos_str?
What do losses_str and sample_infos_str mean in model training ?
| data = tokenizer.decode(data) | ||
| return ds, data, max_loss | ||
|
|
||
| def fetch(self, each_sample_info): |
There was a problem hiding this comment.
Since fetch is user-defined method, then either:
define it in base class as abstract method.
or
the class instance initialization should have a parameter (fetch_func) , which is provided by user.
|
Does this PR still need updates and merging? If so, please reply. This PR will be closed by the end of the month if there is no response. Thanks a lot. |
What changes were proposed in this pull request?
针对loss尖刺的记录和解析提供了一个工具
Why are the changes needed?
为了更多训练的人来使用