IsSynth-Cars / README.md
DanielNguyen121's picture
Expose and verify final Cars test accuracy without rerunning evaluation
440ab7c verified
|
Raw History Blame Contribute Delete
1.06 kB
---
tags:
- image-classification
- research
---
# IsSynth-Cars
## Final Stanford Cars result
**Test top-1 accuracy: 93.3466% (7,506/8,041 correct).** Test macro-F1: **0.932806**. Best validation accuracy: **93.3170%** at epoch 27 (zero based); final validation accuracy: **92.4586%**.
The 40-epoch run used validation to select the best checkpoint, then evaluated that checkpoint on the official 8,041-image test split. This numerical result was recovered and independently verified from the saved predictions and confusion matrix; no additional training or test pass was performed for this reporting update.
See [`reports/final_test_report.md`](reports/final_test_report.md), [`reports/scientific_comparison.md`](reports/scientific_comparison.md), [`configs/cars_baseline.yaml`](configs/cars_baseline.yaml), [`manifests/`](manifests/), and [`runs/cars_issynth_20261025_seed0_support1/`](runs/cars_issynth_20261025_seed0_support1/) for methodology, checkpoints, predictions and reproducibility. Original Stanford Cars image bytes are not redistributed.