dvorak

[22:00] <Kal|Linux> things to learn tomorrow:
[22:00] <Kal|Linux> how do you train the model and save its state
[22:00] <Kal|Linux> without having to train it every time the program is run
[22:00] <Kal|Linux> also how to use the model to predict
[22:00] <Kal|Linux> rather than just saying 'oh i'm 99% accurate'
[22:01] <Kal|Linux> on test cases where the answer may not be "known"
[22:01] <Kal|Linux> and to print out that answer
[22:01] <Kal|Linux> then i can begin constructing the neural network to analyze satellite imagery
[22:02] <Kal|Linux> most of which is going to be finding the imagery

[22:03] <Kal|Linux> hm
[22:03] <Kal|Linux> does anyone know that one website
[22:03] <Kal|Linux> where YE gets lots of old archived dvorak from for EPAC storms
[22:03] <Kal|Linux> ir_bd imagery
[22:04] <+JD|cloud> nrlmry?
[22:05] <Kal|Linux> no like
[22:05] <Kal|Linux> storms from the 80s
[22:05] <Kal|Linux> nrl only goes back to like 95 or something
[22:05] <Kal|Linux> also i wonder if for the source imagery i should use dvorak or normal grayscale ir imagery
[22:06] <Kal|Linux> dvorak might run faster because the images could theoretically be compressed and the neural network could do edge detection
[22:06] <Kal|Linux> but grayscale might have more information to work with
[22:06] <Kal|Linux> and therefore be more accurate?
[22:06] <Kal|Linux> idk
[22:07] <Kal|Linux> i'll do science to it
[22:07] <Kal|Linux> i'll try both
[22:08] * JD|cloud sees that you're not having faith in a linear classifier-based approach
[22:08] <Kal|Linux> what is a linear classifier based approach
[22:08] <+JD|cloud> e.g. the perceptron algorithm with appropriate feature weights
[22:08] <+JD|cloud> with the different classifications being different 5-kt intervals, etc.
[22:09] <+JD|cloud> but stick to a neural network I'd say
[22:09] <Kal|Linux> idk what perceptron is
[22:09] <Kal|Linux> but i also have no idea how neural networks really work either so lol
[22:09] <+JD|cloud> read up on it
[22:10] <Kal|Linux> maybe i'll end up trying multiple approaches
[22:10] <Kal|Linux> then i can see what does best :P

https://www.tensorflow.org/versions/r0.8/tutorials/mnist/pros/index.html#a1

perceptron... other ways of doing it idk

i want it to compare with bins or w/e you know and print out like

35 0.002
40 0.020
45 0.189
50 0.354
55 0.207
60 ...
65
70
75

etc