Tweetovi

Blokirali ste korisnika/cu @ColeJ71909899

Jeste li sigurni da želite vidjeti te tweetove? Time nećete deblokirati korisnika/cu @ColeJ71909899

  1. Prikvačeni tweet
    4. sij

    Finally got around to my last po- I mean _first_ post of the year! AY-3-8500 MiSTer core is released, plus a bunch of other cool silicon reverse-engineering stuff.

    Poništi
  2. proslijedio/la je Tweet
    29. sij

    Posted without comment

    Prikaži ovu nit
    Poništi
  3. 30. sij

    Darn it, he forgot to turn the flash off! 😾

    Poništi
  4. proslijedio/la je Tweet
    29. sij

    I made a goose that destroys your computer Download it free here:

    Prikaži ovu nit
    Poništi
  5. 28. sij
    Poništi
  6. proslijedio/la je Tweet
    25. sij

    --- PSX on FPGA --- New Update #4 for Patrons, Project Update #3 gone public for all : (Please RT)

    Poništi
  7. proslijedio/la je Tweet
    20. sij

    I've written a new article covering the final issue preventing near-perfect emulation of the SNES: exact cycle-timing of the SNES PPUs (video generators.) I've also proposed solutions that are beyond my abilities, so this article is also a call for help.

    Prikaži ovu nit
    Poništi
  8. 24. sij

    After smashing a few bugs, the network generates output again. Keras says the accuracy is over 97%, and there is a very noticeable improvement. Here's the section I showed earlier, before and after these changes.

    Prikaži ovu nit
    Poništi
  9. 23. sij

    Programming survey! (RT for exposure) The variables passed into a function/method are ____, while the variables received by a function/method are ____. (If your answer is "other" say what it is below)

    Poništi
  10. 23. sij
    Prikaži ovu nit
    Poništi
  11. 23. sij

    Ok, I just finished implementing data augmentation, which is a fancy way of saying "rotating the images in my dataset to increase its size" Strangely, this decreased the accuracy. No idea as to why yet.

    Prikaži ovu nit
    Poništi
  12. 21. sij

    I signed up for a class called "Reducing Overfitting in Neural Networks" It would've taught me helpful things, but I dropped out

    Prikaži ovu nit
    Poništi
  13. 17. sij

    That's plenty of time to type in at least a dozen commands through a USB port. Plus I could've had sensitive things open on the screen. It's certainly not a one-time thing either. What bugs me is I have no idea which piece of software is responsible for this delay in locking. 2/2

    Prikaži ovu nit
    Poništi
  14. 17. sij

    When I opened my PC (Linux-based) a few minutes ago, it took a second or two to go from open desktop to login screen. A terminal was open and I managed to type a space in that short window of time. 1/2

    Prikaži ovu nit
    Poništi
  15. 17. sij

    Comments on the future from 2000 (during the Megahertz wars.) Quite funny to read

    Poništi
  16. 15. sij

    Ok, I just improved accuracy a little. Neural nets work best when the input value is in a limited range (eg 0.0-1.0) I was using 8 bit integers (per pixel) but simply dividing those values by 255 resulted in a "happier" neural network, and a 2% improvement in accuracy!

    Prikaži ovu nit
    Poništi
  17. 15. sij

    So soon the network's output should hopefully show noticeable improvement compared to current screenshots, like the one below.

    Prikaži ovu nit
    Poništi
  18. 15. sij

    Another thing I need to do is to figure out how to get CUDA working, so this can be run much faster on my GPU.

    Prikaži ovu nit
    Poništi
  19. 15. sij

    Second fix is to implement data augmentation, which just means automatically creating rotations & transformations of the training images, which increases the amount of training data with minimal work on my part.

    Prikaži ovu nit
    Poništi
  20. 15. sij

    First, adding dropout layers should reduce overfitting, basically the model is memorizing the training dataset instead of "learning" how to tackle the general problem.

    Prikaži ovu nit
    Poništi
  21. 15. sij

    ...problems which are due to an unrepresentative validation dataset and overfitting respectively. Bad results like these are expected since I had no idea what these issues even were when I created this. Now that I'm more educated I can see two immediate fixes...

    Prikaži ovu nit
    Poništi

Čini se da učitavanje traje već neko vrijeme.

Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.

    Možda bi vam se svidjelo i ovo:

    ·