This page is the bill for this solution. It lists the libraries, the machine, the data and the licence of the weights. The worked example ended on the time one picture takes, and this page puts that time in the full cost.
1. What it needs#
It needs a deep learning framework, PyTorch, and the setup to run it. That is a large dependency for a cell whose simplest answer is a page of arithmetic.
It needs a weights file this project does not own. The file for
facebook/sam2.1-hiera-base-plus is 320 MB. It is fetched once into the
solution's weights/ folder and pinned to a version. Solutions 2, 4 and 6 make
their weights here and can make them again. These weights cannot be made here
at all.
It needs arrangements for the keeper, but far fewer than a network trained from scratch. The keeper learns from a short table of numbers, not from pictures. The default is 12 arrangements. That is 36 pictures across the three stations. The scored run got 208 regions from them, so 208 rows to train 180 small trees. The examiner draws and labels them for free. The calibration needs no extra arrangements, because it is trained in folds of the same table.
Training the keeper takes under a second on the processor. No graphics card is needed. The slow part is running SAM 2 over the arrangements to get the regions. That takes about eight seconds a picture on this machine, so the 36 pictures took 273 seconds. At run time, each picture needs one pass of the picture encoder. That is the largest cost here, but it is still small next to one movement of the arm.
The machine is an Apple computer with built-in graphics. The graphics share memory with the processor. PyTorch uses the graphics through its MPS backend, and falls back to the processor without it. There is no separate graphics card and no CUDA. The shared memory is why a model this size fits.
It needs the shading done right. It is the one place where a careless choice makes every later step worse, and nothing reports it.
The licence is permissive. SAM 2's weights are under Apache 2.0 and download without an account. Solutions 3 and 4 use weights under the AGPL, which solution 3's chapter explains. This advantage does not carry over to SAM 3, the newer version, whose weights are gated and come under their own terms.
The next page judges this solution against the other five.