I will implement and reproduce your machine learning paper in pytorch
About this Gig
Need a machine learning paper turned into working PyTorch code, and want to know whether its results actually hold up? I implement the method from the paper, train it, and compare my numbers to the paper's.
I'm an AI researcher at DGIST (South Korea) working on CNNs, vision transformers and continual learning.
Recent example: I reproduced Deep Mutual Learning (CVPR 2018) on CIFAR-100. The paper reports a gain of +2.20 / +1.76 points, and my reproduction measured +2.22 / +1.60. I also found a typo in the paper's results table.
What you get:
- Clean, commented PyTorch code that follows the paper's equations and algorithm
- Training runs with checkpoints and logs
- A report comparing my results with the paper's tables, including any differences and why they happen
- Exact settings so you can rerun everything
Small papers (CIFAR-scale vision, distillation, continual learning) run on my own GPU. Larger papers: add the Cloud GPU extra and I rent a bigger cloud GPU (e.g. A100), so bigger models and datasets are possible.
Please message me the paper link before ordering so I can confirm scope and compute.
Expertise:
Image processing
•
Feature learning
•
Classification
Programming language:
Python
Tools:
Jupyter Notebook
Frameworks:
Scikit-learn
•
PyTorch
FAQ
What if my results don't match the paper?
You still get an honest report showing where the numbers differ and the likely reasons (missing details, randomness, framework changes). I never adjust numbers to match the paper.
Which compute do you use?
My own RTX 4060 (8 GB GPU memory, 16 GB RAM) for CIFAR-scale work, at no extra cost. For larger papers, add the Cloud GPU extra and I rent an A100-class cloud GPU. I quote the hours up front. You can also give me access to your own GPU server.
Do you do coursework or theses?
No. This service is for researchers, startups and engineers who want to evaluate or build on a published method.
