I will optimize your pytorch model with quantization and onnx export
About this Gig
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.
Programming language:
Python
•
Colab
Tools:
Jupyter Notebook
•
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.
