nawta/jaguar-reid-artifacts

Jaguar Re-Identification - Model Artifacts

Kaggle "Jaguar Re-Identification Challenge" コンペティションの学習済みモデル重み・サブミッション・設定ファイルのバックアップ。

Best Score

Public LB 0.948 (exp004 + exp005 weighted ensemble, w=0.7/0.3)

Contents

Model Weights (weights/)

Experiment	Description	LB Score	File	Size
exp004/000	EVA-02-L-448 + GeM + ArcFace + Jaccard Rerank + Optimal Blend (10ep)	0.946	weights/exp004_0938_optimal_blending/000/best_model.pth	~1.2GB
exp005/000	Pseudo-Labeling fine-tune on exp004 (5ep)	0.937	weights/exp005_pseudo_labeling/000/best_model_finetuned.pth	~1.2GB
exp006/001	DINOv2 ViT-L-518 + GeM + ArcFace (20ep)	0.925	weights/exp006_dinov2/001/best_model.pth	~1.2GB

Submissions (submissions/)

各実験の推論結果 CSV (baseline, TTA, rerank, optimal_blend)。

Configs (configs/)

各実験の Hydra 設定ファイル (config.yaml + exp/*.yaml)。

Docs & Logs (docs/, logs/)

• docs/experiments.md: 全実験結果・知見集約
• docs/TODO.md: タスク管理
• docs/KAGGLE_DIRECTION.md: コンペ固有ワークフロー
• logs/Log_*.md: 日別開発ログ

Experiment Summary

Experiment	Backbone	Epochs	Best Loss	Accuracy	LB
exp002/000	MegaDescriptor-B-224	25	-	-	0.781
exp003/000	EVA-02-L-448	10	-	-	0.921
exp003/001	EVA-02-L-448	20	0.3035	96.78%	0.913
exp004/000	EVA-02-L-448	10	0.1355	98.47%	0.946
exp004/001	EVA-02-L-448	20	0.0481	99.42%	0.945
exp005/000	EVA-02-L-448 (PL)	5	0.0731	99.20%	0.937
exp006/000	DINOv2-L-518	10	0.3982	95.14%	0.897
exp006/001	DINOv2-L-518	20	0.0827	98.94%	0.925
exp007/000	EVA-02-L-448 (LwLR+W+EMA)	10	4.4556	48.94%	-
exp007/001	EVA-02-L-448 (LwLR0.95+W+EMA)	10	0.3466	95.56%	-
exp007/002	EVA-02-L-448 (W+EMA)	10	0.1904	97.83%	-
Ensemble	exp004+exp005 (w0.7/0.3)	-	-	-	0.948

Key Insights

• EVA-02 Large 448px が最も効果的なバックボーン (MegaDescriptor-B-224 比 LB +0.140)
• Jaccard Re-ranking + Optimal Blending (20% raw + 80% reranked) が後処理として有効
• 20ep 延長学習は微小改善〜悪化: exp003は悪化 (0.921→0.913)、exp004はほぼ同等 (0.946→0.945)
• Layer-wise LR Decay は EVA-02 Large に有害: decay_rate=0.95 でも全層同一LR に劣る
• Warmup + EMA は微小劣化: 10ep では warmup の序盤学習遅延を回収しきれない

Reproduction

Code

git clone git@github.com:nawta/Jaguar_Re_Identification.git
cd Jaguar_Re_Identification
git checkout lb-0.948

Environment

# Python environment
uv sync

# or Docker
make build && make bash

Data

Competition data is NOT included. Download from Kaggle:

kaggle competitions download -c jaguar-re-id
unzip jaguar-re-id.zip -d input/

Training

# Best single model (exp004)
uv run python -m experiments.exp004_0938_optimal_blending.run exp=000

# Pseudo-labeling (exp005, requires exp004 weights)
uv run python -m experiments.exp005_pseudo_labeling.run exp=000

# DINOv2 (exp006)
uv run python -m experiments.exp006_dinov2.run exp=000

Ensemble (LB 0.948)

exp004 と exp005 の submission.csv を weighted average (w=0.7/0.3) で結合:

import pandas as pd
df004 = pd.read_csv("output/experiments/exp004_0938_optimal_blending/000/submission.csv")
df005 = pd.read_csv("output/experiments/exp005_pseudo_labeling/000/submission.csv")
df_ens = df004.copy()
df_ens["score"] = 0.7 * df004["score"] + 0.3 * df005["score"]
df_ens.to_csv("submission_ensemble.csv", index=False)

Not Included

• Competition input data (input/, ~17GB): Re-download via kaggle competitions download -c jaguar-re-id
• Inferior model weights: exp002 (LB 0.781), exp003 (LB 0.921), exp007 (exp004 以下)
• Python virtual environment (.venv/): Recreate via uv sync
• wandb logs: Synced to wandb.ai/nawta1998/jaguar-re-identification