This page lists what this solution costs. It covers the libraries, the machine, the data and the licence.
1. What it needs#
This is one of the harder solutions in the set to set up. The list below says what it needs, from the largest cost to the smallest.
- A deep learning framework. It needs PyTorch and the RF-DETR package. That is a large dependency for a cell whose simplest solution is a page of arithmetic.
- A downloaded weights file. The borrowed weights are 128 MB. They are fetched, not stored with the code. This project cannot make them, so they must be trusted.
- A training set. The examiner draws and labels it for free, including the crowded arrangements. This is the cheap part, and it is why fine-tuning makes sense here.
- Training time. The fine-tune is the largest cost. It is far shorter than training from random numbers. The one fine-tune measured here took 932 seconds. Other training runs shared the machine, so that is an upper limit. The second way was stopped before it finished, so it has no time.
- Hardware it fits on. The machine is a laptop. Its graphics processor shares memory with the main processor. PyTorch reaches it through Apple's Metal Performance Shaders, which PyTorch calls its MPS backend. The code uses MPS when it can, and the main processor when it cannot. It prints which one it got. There is no NVIDIA card and no CUDA. Shared memory helps the model fit. If the model is still too large, a smaller size can be chosen.
- Weights kept in step with the cell. Change the camera, the grey shading or the range of a kind, and the weights file is out of date. No code test will notice.
Running the model is cheap. One pass over one picture takes a small part of the seconds an arm movement takes. So almost all of this solution's cost is in building it, not in running it.
The licence is Apache 2.0. It lets you use, change and ship the model with no duty on the code around it. The two YOLO solutions use the AGPL instead, which the licence section of solution 3 explains.
How it compares next weighs these costs against what the solution scores.