Corrosion segmentation - APPENDIX
Report on training experiments to make the corrosion model as light as possible while keeping correct performances.
Created on March 14|Last edited on March 14
Comment
Appendix
Appendix with detailed results
UNET 6.4M - exp 53
Metrics on train set
{'accuracy': 0.9785500764846802, 'precision': 0.8026665449142456, 'recall': 0.9166544675827026, 'dice': 0.8157516121864319, 'iou': 0.744350254535675}
Metrics on validation set
{'accuracy': 0.9747447371482849, 'precision': 0.7753747701644897, 'recall': 0.9004101157188416, 'dice': 0.7825585603713989, 'iou': 0.7083650231361389}
Distillation UNET 576k - exp 1004 :
Metrics on train set
{'accuracy': 0.9769202470779419, 'precision': 0.8295949697494507, 'recall': 0.8614419102668762, 'dice': 0.8057383894920349, 'iou': 0.7301884889602661}
Metrics on validation set
{'accuracy': 0.9750799536705017, 'precision': 0.8122709393501282, 'recall': 0.8545907139778137, 'dice': 0.7893723845481873, 'iou': 0.7154967784881592}
Distillation UNET 576k - exp 1004 :
Distillation UNET 576k - exp 1004 :
1004 - UNET 576k -2.2Mb on disk
{'accuracy': 0.9750800728797913,
'precision': 0.8122708797454834,
'recall': 0.8545903563499451,
'dice': 0.7893725037574768,
'iou': 0.7154967784881592},
1004 - UNET 576k + Weights QUANTIZED 8 bits ~ 550kb on disk
{'accuracy': 0.975077748298645,
'precision': 0.8139002919197083,
'recall': 0.8537895679473877,
'dice': 0.7895622849464417,
'iou': 0.7158442735671997},
1004 - UNET 576k + Weights QUANTIZED 4 bits ~ 275kb on disk
{'accuracy': 0.9681149125099182,
'precision': 0.7797178030014038,
'recall': 0.783839225769043,
'dice': 0.7341292500495911,
'iou': 0.6515549421310425},
Add a comment