I will optimize your pytorch model with quantization and onnx export
Computer Vision and AR Engineer, PhD Researcher
Über diesen Service
Is your PyTorch model too slow, too large or too expensive to run in production? I am a PhD researcher in computer vision with peer-reviewed publications on model compression and knowledge distillation. I will make your model smaller and faster while keeping accuracy under control.
What I can do:
- Post-training quantization (FP16 / INT8) and quantization-aware training
- Structured and unstructured pruning
- Knowledge distillation into a smaller student network
- Export to ONNX / ONNX Runtime, plus TensorRT or Core ML where your hardware supports it
- Before/after benchmarks: latency, model size, memory and accuracy
What you get:
- Optimized model files and clean, documented Python code
- A short report explaining every change and its accuracy cost
Please message me before ordering with your model architecture, framework version, target hardware (GPU, CPU, Jetson, mobile) and accuracy budget, so I can confirm the right package.
Programmiersprache:
Python
•
Colab
Tools:
Jupyter-Notizbuch
•
opencv
•
PyTorch
Frameworks:
keras
•
PyTorch
FAQ
Will optimization reduce my model's accuracy?
Some techniques can cost a small amount of accuracy. We agree on an accuracy budget before I start, I measure every step and you get before/after numbers, so there are no surprises.
Which models and frameworks do you support?
Mainly PyTorch vision models: classification, detection, segmentation and similar tasks. TensorFlow/Keras models can often be handled through ONNX. If your model is unusual, message me first and I will tell you honestly whether I can help.
Do you need my training data?
Quantization and ONNX export only need a small calibration/validation set. Distillation and fine-tuning need access to training data. Your model, code and data stay confidential and are never reused.
Which hardware do you test on?
I benchmark on NVIDIA RTX GPUs (TensorRT, ONNX Runtime GPU), x86 CPUs, Apple Silicon Macs and Android phones. For Jetson or other edge boards I deliver the exported model plus a benchmark script you can run, or I test with access you provide.
What exactly do I receive?
The optimized model files (for example ONNX or a TensorRT engine), clean Python code that reproduces every step, and a short report with before/after latency, model size and accuracy.
