Closed ivanzvonkov closed 1 year ago
Summary about the Malawi map quality from the call (09/27/2022)
Next steps
Got points from Blake will wait for #251 to finish
So far, I have received 28 corrective label files from Ivan. Tried adding the new labels but the Pull Request failed because of the unresolved billing issue
This is an excellent report @MsPixels!! Thank you for doing this analysis. Did you share it with Christina?
Thanks Ivan. Yep, I shared this with Christina
@ivanzvonkov, this is the error I get when I run the model
The Malawi model performs worse with all the new data than when only using high quality data (2 or more labelers). See comparison here: https://api.wandb.ai/links/nasa-harvest/8b6sizbh
This indicates that datasets such as MalawiCorrectiveLabels2020
may still contain incorrect labels that confuse the model and makes #164 higher priority.
Validation metrics are similar to what they were before corrective labeling so I don't think making another map is necessary until there are better metrics. I am thinking of moving this issue to Paused
status and reassessing priority on Monday. @MsPixels @hannah-rae let me know what you think.
On second thought, the test set metrics of the Malawi model trained on high quality data are higher than I've seen before (higher than validation set even), so maybe we should make a map: https://github.com/nasaharvest/crop-mask/blob/c58a4426d5583d87eb48ab807c4a115437c9bb86/data/models.json#L70
After making the map, I got these results - Comparing V1 and V2. There's definitely some changes
Potentially do post processing (E.g. filtering out fallow fields re Ben's work, sieve for cleaning noisy predictions [window])
Start year: 2020 Start month: September