Well, there are tons of tutorials say https://keras.io/examples/vision/mixup/ : this examples can be trained ... for 30 * 30 0/1 coded digits images in a reasonable time.So instead of trying to find something that might be of use, how do we train our own AI's on a Pi5.
Can it even be done?
ca 20% of https://keras.io/examples/vision/nice_i ... _tutorials cannot run on a PI5, because some library could not be installed (python library management is chaotic)
On the remaining, 60/70 % cannot run because of RAM shortages (a Pi5 with 8G is very small; same issue occured with a Pi4: when tutorials can run, they are ...3 times faster/less slow on a 5)
Image recognition is very popular and one can find tons of pretrained DNNs; most of them can run in inference mode on a Rpi x (x==4 is 2/3 times slower than x==5, once DNN is loaded into RAM)
Perhaps training only one layer, according to a new problem, migh be do able (on a PI3, it took ... days or weeks; on Pi5, it may tajke hours or days...) https://www.tensorflow.org/guide/keras/ ... ning?hl=en Just evaluating them -if one has a new, good quality (image+annotations) database- is very time consuming (format conversion...)
Statistics: Posted by dbrion1 — Mon Apr 08, 2024 2:15 pm
