July Logs

01

Control - invoke conscious thinking by repeating so gain a habit of subconsciously being a consious thinker to greater productivity.

02

#recall geez, this feeling of exploring direction that dont know if an answer exists is so familar, perhaps that one year project isnt that meaningless from some angle.

04

#NN When in double, use adam not SGD

05

love code file with a test case at the end so you know how the code work immediately

learn how to google - good keyword, etc - thats all

#reset 4 - ok, really high oscillation in productivity after final result released, happens almost every time. Hope this is the last time.

07

Responsibility in team work and contingency.

nowadays, could just judge the quality of a course by comparing with its online e.g. coursera counterpart, unfortuntely recently have taken one that is just disappointing, master degree, sometimes, rather, money revenue for uni.\

#nn Ian Goodfellow move back to google from apple.

10

#debug emm, get ‘RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.FloatTensor) should be the same’ and it turns out just forget to add .cuda() on model since data and model cannot be on different device.

11

# 雪中悍刀行， a novel I started reading at year 8 around, after that, I listen it twice until the 13rd one and re-read it twice until the 13rd one with a gap of 1 year or 2, since 13rd series is the last chapter. I.e. a novel read four times without finishing it, very indeed the first time, another book 傅雷家书, read first 20 pages and give up on a plane back sydney from shenzhen becaues it is too relatable, every sentence in that series of letter exchange just recall memory.

12

#uni #rant no RL, no CV (deep learning not image processing), no GAN, no auto driving, just want to dive deeper in AI, only sees these courses in other country..

14

#research just about to thought the research is not working again, a obvious bug lead to me and my supervisor helping me debug, then find out 2 more logic / theory bugs that is not obvious, hope no major bugs any more and can get the reliable data finally. pews japan vlog is way to chill.

another reason why cuda run out of memory could be cached effect from the jupyter notebook global variables

#solution about to set up ali cloud for autodl becaues the download speed is too slow but then realise it is rather more ppl using the internet at this apartment and so I simply switch to phone hotspot just for downloading the model, problem solved.

15

#network autodl server is in china, so today, sort of expectingly, when download cifar10 hosted in torontoU, its socket disconnected. #juan saw her in the online study room whole night yesterday in my timezone, continue at 8:30am today, stil saw her.

#research 2 tiems bug + 2 time implementation understanding mistakes - rerun experiment 4 times now.

#uni med students seems to be quite some more work, quite some in the online study gp

16

#mistake forget to turn off GPU instance yesterday, oh, no, now I love colab now.

#ai one thought on AI development thou, “Be Fearful When Others Are Greedy”, maybe too many rn, should prob get into sth more specific and valued.

online study gp - 1mill neets, A level students, btw, just really want to say to those: “health is weath”… #quote “not come here to take part, I come here to take over”

18

NVIDIA @ptrblck_de - knew so much, almost the answer all important soln for pytorch Forum.

“windowless laboratory “ that hits.

21

four eggs, two tomato, one instant noodle, 15mins, problem solved. It turns out this research is exactly what I looking for when I initially thinking on it - really helps me inspect every params in a model and understand in terms of values not taught concept.

22

#colab #nn colab zip file like e.g. tar file when upload to the running environment will be corrupted - second times, 2 mins wasted.

#code “dependency hell” I can see it easily

jupyter book build static web needs .nojeykll for some reason it disappear and the website becomes a skeleton, 18mins debug get it back, hooo.

23

#rant My window os reboot without any notice again, no wonder so much ram usage but not much shown in the task manager, it is all come from the window update, window os - make u feel no control over ur pc, i conclude it worse than templateOS, worse ever in terms of ux.

24

#nn large batch size could be use to see the approx how well the model could learn, e.g. 512 for cifar10 instead 64, if it learn in 512, it will for 64 as well.

26

#retrospect think a mill times before say

29

night run is addictive.

30

School starts next monday