PyTorch → ONNX → FastAPI → Docker. FER2013 baseline + robustness + Kalman smoothing + ONNXRuntime API.
A modular perception pipeline designed for robust inference under noisy conditions. The project demonstrates end-to-end ML deployment workflows, from ONNX model inference to containerized API services, with Kalman-based post-processing for stability.
-
Model Inference
- Exported PyTorch model to ONNX and executed inference via ONNX Runtime.
- Simulated noisy inputs (Gaussian blur, occlusion, variance shifts).
-
Post-processing
- Applied Kalman smoothing to stabilize predictions under degraded conditions.
- Quantified robustness improvements with spectral/variance analysis and condition number regularization.
-
Deployment
- Containerized services with Docker (CPU/GPU modes).
- Lightweight REST API using FastAPI for easy evaluation.
-
Profiling & Stability
- Latency and resource profiling for inference pipelines.
- Robust accuracy curves under varying noise levels.
We tested the baseline CNN model under several perturbations. Dataset: FER2013 (or FakeData for pipeline validation).
| Scenario | Accuracy |
|---|---|
| clean | 0.51 |
| gaussian_0.10 | 0.40 |
| blur_k5 | 0.45 |
| occlusion_25 | 0.30 |
| bright_1.4 | 0.49 |
Applied a simple 1D Kalman filter on the class probability sequence.
- Raw accuracy: 0.51
- Smoothed accuracy: 0.74
Exported PyTorch model to ONNX:
- File:
models/fer2013_cnn.onnx - Verified with onnxruntime (max error < 1e-5)
Test inference results saved in:
results/infer_onnx.csv
pip install -r requirements.txt
uvicorn src.train.infer_api:app --host 0.0.0.0 --port 8080
# curl -F "file=@/path/to/img.jpg" http://127.0.0.1:8080/infer_form