So Categorical Cross Entropy is a nightmare! Pure nightmare! However, with some scotch and tape, it seems to now be working. Fiiiiinally.
Yeah, yeah, "RMSE AS A METRIC FOR MULTI CLASS CLASSIFIER!?", but why are you looking to the right anyway, pls enjoy the CategoricalCrossEntropy Loss on the left. I'll write a proper metric next.
Finished up with the second workweek. I feel like this one went much much better. Got assigned a new client, all mine this time. Did some data science -mostly cleaning, but also feature generation and model tuning. It's kinda interesting since I'm working with really tiny datasets, unlike what I'm used to, but even with them I am able to create decent linreg models. On a different note, wasted hours of my weekend playing this stupid! io game which I won't link cause of how addictive it is. Finally, watched Lex's podcast where the guest said Python is bad for some things like ML framework. raML sigh Evening update: spent a lot of time trying to implement softmax. Still not finished.
The first working week ended very well. Since I didn't like my productivity level, I compensated for low productivity with a higher number of hours, i.e. since Satisfaction = Productivity * Hours, one can reach Satisfaction at lower Productivity by increasing the number of Hours. Actually, there is more to say about these variables: S, P, H. As you progress, your P steadily increases, and therefore really is a function of Hours, or to be more exact the function of Total Hours with some time decay parameter (i.e. if you stop putting hours, P decays). If you periodically fall short of Satisfaction, that's what we call on the internet - GG. The only variable you can really control is H, which also has a soft lower bound of 40hr/wk. The upper bound however also exists, and in fact, if you put in 20hours one day, the next day you might only be able to put in 2 due to exhaustion. I think a reasonable upper bound is somewhere around 10-12 hours. There are tricks to stretch this upper bound by putting in hour towards some other thing, like a hobby, so both limits are somewhat flexible.
Wasn't really happy with the work progress today, putting in extra hours to get to the satisfactory point.
TLDR of the first working day: tech stack = hard, math = makes sense. Setting everything up was pretty exhausting, but I'm pretty sure, if I haven't missed anything (too) important, I'm basically close to being ready to do Data Science, woah. Another awesome thing happened - my mom got into the MRI program at a really good school!! Really happy for her. She didn't think it was possible because of how much bureaucratic push back she got on her application (not finishing college in the home country, etc.) but she got in! That's really awesome, hopefully, the program will be worth it.
