Kyle Kastner

@kastnerkyle

computers and music are fun. music:

Montréal
যোগদান করেছেন জানুয়ারী ২০১১

মিডিয়া

  1. ১৭ জুলাই

    Another session I really enjoyed was this talk on "Deep Exploration" from in the PGMRL workshop, found a similar version of the talk from DALI online

  2. ১৫ জুলাই

    Excited about Anna Huang (et. al. w/ Magenta)'s work with a modified Transformer for generating piano roll-likes, a) works stunningly well, even on small corpora like Bach, and b) is published in workshop form! PAPER: && AUDIO:

  3. ১৩ জুলাই

    At that point, you may be better off doing pure phonemes + splitting SIL based on duration, and using grapheme-to-phoneme (g2p) tools to handle the fact that few general users like typing text on a per-phoneme basis

    এই থ্রেডটি দেখান
  4. ৫ জুলাই
    -কে উত্তর দিচ্ছেন

    My AC doesn't really work, this is basically my status. Having memories of the lab before a deadline, and not in a good way...

  5. ৫ জুলাই

    Reminded of this great video about face lighting/orientation/pose

  6. ২৭ জুন
    -কে উত্তর দিচ্ছেন

    "deep X" is fundamentally a Texan endeavor after all ...

  7. ২৫ জুন

    My current plan for getting the university to give me a PhD

  8. ২৪ জুন
    -কে উত্তর দিচ্ছেন

    In one day you can see potential, but not that Charles Atlas Seal of Approval

  9. ২৪ জুন

    but 14 days is better

    এই থ্রেডটি দেখান
  10. ২৪ জুন

    My mental rule for training time of neural TTS - "in just 7 days, I can make you a [TTS]"

    এই থ্রেডটি দেখান
  11. ২৩ জুন

    ... nothing else to say than R I P

    এই থ্রেডটি দেখান
  12. ২৩ জুন

    and he kept doing it, year after year, song after song. His drums always pushed the song forward, while complimenting the other instruments at the same time.

    এই থ্রেডটি দেখান
  13. ২৩ জুন

    groovy, heavy, amazing. This song changed my thoughts on what metal music could sound like... and the drums are the heart of it all. RIP

    এই থ্রেডটি দেখান
  14. ২০ জুন

    When the TTS model starts "talking back" after training overnight, and the loss is still going down

  15. ৫ জুন

    Interesting behavior in training on this task - you can almost *see* where it (probably?) learned an algorithmic solution. This is very early on in training, about 1000 minibatches of size 64. Loss tends to spike once it is very low, consider grad clip as well

    এই থ্রেডটি দেখান
  16. ৫ জুন

    Ever want a Rube Goldberg-ian method for reversing words? Try Transformer (Attention Is All You Need) - it does a really nice job with this. And reversing words with neural nets without mistakes is not as easy as it sounds (my goto test for seq2seq)...

    এই থ্রেডটি দেখান
  17. ১ জুন

    "Geodesics in Heat" + pathfinding on regular 2D grid ala , using 2 different integration times for heat diffusion. Longer integration times == more "natural / cautious" paths, see Figure 12

  18. ১ জুন
    -কে উত্তর দিচ্ছেন

    I confirm it can hold references to *at least* 596 papers

  19. ২৯ মে

    This talk makes me want to study differential geometry . Lots of visualizations and intuition about complex topics, with a little math on top to tie it all together.

  20. ২৯ মে

    If you actually care about the distances, it looks like this. But "prism" is a lot better than "viridis" for that funhouse/carnival vibe

    এই থ্রেডটি দেখান

লোড হতে বেশ কিছুক্ষণ সময় নিচ্ছে।

টুইটার তার ক্ষমতার বাইরে চলে গেছে বা কোনো সাময়িক সমস্যার সম্মুখীন হয়েছে আবার চেষ্টা করুন বা আরও তথ্যের জন্য টুইটারের স্থিতি দেখুন।

    আপনিও পছন্দ করতে পারেন

    ·