4/12/2021 Gregorian Calendar 16:11:56 HST
This day I spent the whole time trying to learn python to do the solar panel counting AI. I have MNIST ai working, but the way it loads the data is through the keras which puts it into a numpy array and everything for you. If I wanted to make an AI that takes an image, I would have to do something similar. The MNIST thing is in a 60000x1x28x28 numpy array (the 60 thousand is how many images) or whatever its called, but I had to first figure out how to allocate an array with that structure but mine would be like 60000x1x128x128x3 (3 because RGB) or whatever size the image is going to be (and probably not 60000, more like 100 since there isn't that much data. Python doesn't do memory stuff the same way and I usually worst case in C is to just manually push the memory into the array, but python can't do individual register shennanagins, so I didn't know how to put an image into an array format.
4/14/2021 Gregorian Calendar 15:29:06 HST
Today I fixed the wierd bug on the blockmatching like 5 weeks later. It was just the threading overlapped because in the for loop it counted up to the height of the image divided by the blocksize, but if it isn't a whole number it would return a float, but I was accidentally saving it as an integer and it would trunkate then in the for loop allocating the threads would make the last 2 collide and overwrite each others output messing up the memory in the array which when displayed would make it look like a white pixel instead of red since it is just messed up.
So this week I failed to do an AI because python is weird then messed around with multi-threading until I fixed a minor bug in a previous program that would also be the input into an AI.
bruce_a
Comments