0.3884
Model Evaluation
Model Evaluation, Data Leakage Analysis, Error Analysis on BDD Dataset
Created on March 7|Last edited on March 7
Comment
On this report, we validate the data splitting to see how the data leakage going. In the script, we split the data using StratifiedGroupKFold to make sure the data is distributed based on rare class.
This is the model metrics that are summarized based on various runs.
Model IoU
Run set
50
This is the evaluation metrics on the validated model run
The registered model can be accessed via gustiwinata/model-registry/BDD Semantic Segmentation:v0
Run set
50
The model is needed to be focused on train on rare classes like Person, traffic light and bicycle since the model evaluation are scored poorly on those classes. The sweep efforts is already been applied, but since for sweeping on batch_size is already taking a whole day and I did not applied models arch sweeping parameters, yield the model to not applied to such classes. This could be averted by training the data in a better machine and run sweeps on multiple models arch.
Here are the run plots summary from various runs for references.
Run set
50
Add a comment
Gusti Triandi Winata •
The model is needed to be focused on train on rare classes like Person, traffic light and bicycle since the model evaluation are scored poorly on those classes. The sweep efforts is already been applied, but since for sweeping on batch_size is already taking a whole day and I did not applied models arch sweeping parameters, yield the model to not applied to such classes. This could be averted by training the data in a better machine and run sweeps on multiple models arch.
maybe should add more details
Reply