Update Gyuyig vs Tsugdri binary classifier: center_crop 224 metrics, confusion matrix, training history

92d18ee verified 5 days ago

4.11 kB

language:
  - bo
license: apache-2.0
tags:
  - image-classification
  - tibetan
  - script-classification
  - dinov3
  - binary
library_name: transformers
pipeline_tag: image-classification
base_model: facebook/dinov3-vits16-pretrain-lvd1689m
datasets:
  - BDRC/gyuyig-tsugdri-binary-balanced-script-classification-dataset
metrics:
  - f1
  - accuracy
  - auc

Gyuyig vs Tsugdri Binary Script Classifier (DINOv3 ViT-S)

Fine-tuned DINOv3 ViT-S for parent script classification:

Gyuyig, Tsugdri

Experiment: dinov3_gyuyig_tsugdri_binary (gyuyig_tsugdri_binary_classification)
Pooling: ViT CLS token (last_hidden_state[:, 0, :])
Weights: final_model.pt (best validation macro-F1 across stages A/B/C)

Warm-start: BDRC/4-class-balanced-script-classifier (final_model.pt — prior test acc 92.1%, macro-F1 0.921)

Data

Split	Source
Train / val / test	BDRC/gyuyig-tsugdri-binary-balanced-script-classification-dataset

Test split: balanced benchmark (60 images per parent class, held out of training).

Preprocessing

Split	Mode	Size
train	`center_crop`	224
val	`center_crop`	224
test	`center_crop`	224

Validation metrics (n=60)

Metric	Value
Accuracy	91.7%
Macro F1	0.916
Weighted F1	0.916
AUC-ROC	0.931
Loss	0.3915

Best checkpoint: best_stage_c_last_blocks.pt epoch 1 val macro-F1 0.916

Per-class (validation)

precision    recall  f1-score   support

      Gyuyig       0.88      0.97      0.92        30
     Tsugdri       0.96      0.87      0.91        30

    accuracy                           0.92        60
   macro avg       0.92      0.92      0.92        60
weighted avg       0.92      0.92      0.92        60

Test / benchmark metrics (n=120)

Metric	Value
Accuracy	85.0%
Macro F1	0.848
Weighted F1	0.848
AUC-ROC	0.930
Loss	0.4047

Per-class (test)

precision    recall  f1-score   support

      Gyuyig       0.78      0.97      0.87        60
     Tsugdri       0.96      0.73      0.83        60

    accuracy                           0.85       120
   macro avg       0.87      0.85      0.85       120
weighted avg       0.87      0.85      0.85       120

Training

Stage	Epochs	LR head	LR backbone	Unfrozen blocks
A	7	0.0005	—	0
B	10	0.0001	1e-05	4
C	12	5e-05	1.5e-05	8

Setting	Value
Scheduler	`cosine_warmup`
Class weights	`custom`
Label smoothing	0.05
Dropout	0.1

Confusion matrix (test)

True \ Pred	Gyuyig	Tsugdri
Gyuyig	58	2
Tsugdri	16	44

Files

File	Description
`final_model.pt`	Best val-F1 weights + label maps
`results.json`	Full metrics, history, warm-start info
`config.yaml`	Training config
`model_card.json`	Summary metadata
`confusion_matrix.json` / `.png`	Test CM
`training_history.png`	Stage loss / val F1 curves
`split_stats.json` / `.md`	Per-class split counts
`inference.py`	Classify image paths
`requirements-inference.txt`	Pip deps

Inference

pip install -r requirements-inference.txt
python inference.py --checkpoint final_model.pt --image path/to/page.jpg --preprocess resize_letterbox --preprocess-size 224

Reproduce training

python experiments/gyuyig-tsugdri/train.py --config experiments/gyuyig-tsugdri/config_warmstart.yaml

Model repo: BDRC/gyuyig-tsugdri-binary-script-classifier