Release time distributed model trained on ~250k videos #226
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
✅ Deploy Preview for silly-keller-664934 ready!
To edit notification comments on pull requests, go to your Netlify site settings. |
Codecov Report
@@ Coverage Diff @@
## master #226 +/- ##
======================================
Coverage 87.0% 87.0%
======================================
Files 29 29
Lines 1937 1937
======================================
Hits 1686 1686
Misses 251 251
|
|
🚀 !!! |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Replaces time distributed model trained on 15k videos with one trained on 250k videos. The same (improved) frame selection is used for both).
Key decision
This removes the 15k model so there is only one "time distributed" model available as an official model.
Rationale
As seen in the
val_metrics.json, the new model does not have improved performance on every species judged by the F1 score. However, the new model has been evaluated on 30k instead of 3k videos and includes a greater diversity of videos given the inclusion of the Chimp & See data. From exploration into calibration and gains curves, it appears that there is still signal for the species with low or zero F1 score, it is just that probabilities are of lower confidence (e.g. highest probabilities still contain an animal, but those probabilities are less than 0.5). The old model was also trained on a much more balanced dataset (blanks at 15%) compared to the new model (blanks account for around 40%). The new model does significantly better on blank performance (f1 of 0.83 vs. 0.54) while also maintaining high performance on other species.Closes https://github.com/drivendataorg/pjmf-zamba/issues/129