Deep Learning Model Files
In the deep learning lifecycle, different model formats serve different purposes, ranging from training to deployment.
Format Relationships
The diagram below illustrates the standard conversion chain for modern deep learning models.
Supported Formats
| Format | Role | Primary Use |
|---|---|---|
| .pt / .pth | Training | Native PyTorch storage for weights and architectures. |
| .onnx | Conversion | Universal bridge between frameworks (Open Neural Network Exchange). |
| .engine / .plan | Deployment | High-performance optimized inference on NVIDIA GPUs. |
| .wts | Bridge | Plain text weight storage for custom TensorRT builders. |
Recommendation
The standard path for 90% of production models is: .pt → .onnx → .engine.
Key Concepts
- Framework Independence: ONNX allows you to move models between different tools without framework lock-in.
- Hardware Optimization: TensorRT engines are bound to specific GPU architectures and driver versions to achieve peak performance.
- Conversion is One-Way: While you can go from
.ptto.engine, you generally cannot reverse the process.