← Back to Home

Artifact Reference

Every run generates a complete artifact suite for total traceability. Both training-only and optimization workflows share the same RunPath orchestrator, producing BLAKE2b-hashed timestamped directories.

Directory Structure

Standard Training Run

outputs/20260123_bloodmnist_minicnn_b4a8f1/
├── figures/
│   ├── confusion_matrix.png           # Per-class prediction matrix
│   ├── training_curves.png            # Loss/AUC over epochs
│   ├── training_curves.npz            # Raw metrics data for custom plots
│   └── sample_grid.png                # Augmentation preview (12 samples)
├── reports/
│   ├── training_summary.xlsx          # Complete metrics spreadsheet
│   ├── config_snapshot.yaml           # Frozen config at run start
│   └── requirements.txt              # Pinned pip freeze at run start
├── checkpoints/
│   └── best_mini_cnn.pth              # Best model weights (by val AUC)
├── exports/
│   └── model.onnx                     # Production-ready ONNX export
└── logs/
    └── orchestrator_YYYYMMDD_HHMMSS.log

Optimization Run

Optimization runs include all standard artifacts plus additional analysis files:

outputs/20260123_organcmnist_efficientnetb0_a3f7c2/
├── figures/
│   ├── confusion_matrix.png           # Final model predictions
│   ├── training_curves.png            # Best trial training curves
│   ├── training_curves.npz            # Raw metrics data
│   ├── sample_grid.png                # Augmentation preview
│   ├── param_importances.html         # Interactive importance plot
│   ├── optimization_history.html      # Trial progression over time
│   ├── slice.html                     # 1D parameter effect analysis
│   └── parallel_coordinate.html       # Multi-dimensional parameter view
├── reports/
│   ├── training_summary.xlsx          # Best trial metrics
│   ├── config_snapshot.yaml           # Initial config
│   ├── requirements.txt              # Pinned pip freeze at run start
│   ├── best_config.yaml               # Optimized hyperparameters
│   ├── study_summary.json             # All trials metadata
│   └── top_10_trials.xlsx             # Best configurations ranked
├── checkpoints/
│   └── best_efficientnet_b0.pth       # Best model weights
├── exports/
│   └── model.onnx                     # Production export
├── database/
│   └── study.db                       # SQLite storage for resumption
└── logs/
    └── orchestrator_YYYYMMDD_HHMMSS.log

Artifact Details

Figures

File	Description	Generated By
confusion_matrix.png	Per-class prediction heatmap showing true vs predicted labels	Training, Optimization
training_curves.png	Loss and AUC metrics plotted over epochs	Training, Optimization
training_curves.npz	NumPy archive with raw curve data for custom visualization	Training, Optimization
sample_grid.png	12-sample grid showing augmentation effects	Training, Optimization
param_importances.html	Interactive Plotly chart showing hyperparameter importance	Optimization only
optimization_history.html	Trial progression showing objective value over time	Optimization only
slice.html	1D parameter slice plots showing individual parameter effects	Optimization only
parallel_coordinate.html	Multi-dimensional view of parameter relationships	Optimization only

Reports

File	Description	Generated By
config_snapshot.yaml	Frozen configuration at run start (immutable record)	Training, Optimization
requirements.txt	Pinned pip freeze output capturing exact dependency versions	Training, Optimization
training_summary.xlsx	Excel spreadsheet with metrics, predictions, and class-wise breakdown	Training, Optimization
best_config.yaml	Optimized hyperparameters ready for production training	Optimization only
study_summary.json	Complete study metadata including all trial results	Optimization only
top_10_trials.xlsx	Top 10 configurations ranked by objective value	Optimization only

Models & Exports

File	Description	Generated By
best_<arch>.pth	PyTorch checkpoint with best model weights (selected by validation AUC)	Training, Optimization
model.onnx	ONNX export for production deployment (opset 18, dynamic batch)	Training, Optimization

Database

File	Description	Generated By
study.db	SQLite database storing Optuna study state for resumption	Optimization only

Sample Artifacts

Explore real experiment outputs in the artifacts directory:

• Excel Reports: Training metrics, predictions, class-wise analysis
• YAML Configs: Frozen configurations and optimized hyperparameters
• HTML Visualizations: Interactive Plotly charts for optimization analysis

---

← Back to Home | Optimization Guide | Export Guide