Ikibazo gihuza ibibazo · Hagati Urwego

Wasserstein GAN Ikibazo

Kugenzura intera ya Wasserstein, amanota yo kunegura, imiterere ya Lipschitz, kugabanya ibiro, ibihano bya gradient hamwe na GAN imyitozo itajegajega.

Inzira zijyanye no kuyoboraWASerstein GAN
Ikibazo 1 Bya 8

Nibihe bipimo WGAN ikoresha kugirango igereranye kugabanwa kwukuri kandi kubyara?

Reba buri kibazo8 ibibazo

Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.

  1. Nibihe bipimo WGAN ikoresha kugirango igereranye kugabanwa kwukuri kandi kubyara?
    • Intera ya Wasserstein (isi-yimuka)
    • Kullback-Leibler gutandukana
    • Cosine
    • Intera ya Euclidean
  2. Muri WGAN, umuyoboro wari ivangura wahinduwe iki, kandi kuki?
    • 'Umucamanza' kuko itora yego cyangwa oya
    • 'Kunegura' kuko itanga amanota ahoraho aho kuba bishoboka
    • A 'sampler' kuko itanga urusaku
    • 'Akayunguruzo' kuko gakuraho amashusho mabi
  3. Kuki abanenga WGAN bagomba guhagarikwa kuba 1-Lipschitz?
    • Kugirango itoze byihuse kuri GPUs
    • Kugabanya umubare wibipimo
    • Kuberako Kantorovich-Rubinstein ikoreshwa mu kubara intera ya Wasserstein isaba gradients imipaka;
    • Guhatira gusohora integer
  4. Nigute WGAN yumwimerere yashyize mubikorwa imbogamizi ya Lipschitz?
    • Wongeyeho abata ishuri
    • Ukoresheje icyiciro gisanzwe gusa
    • Mugukonjesha amashanyarazi
    • Mugukata uburemere bw'abanenga kurwego ruto
  5. Ni ikihe kibazo WGAN igabanya cyane ugereranije na vanilla GAN?
    • Uburyo bwo gusenyuka no gutakaza amakuru
    • Birenzeho kubizamini
    • Buhoro buhoro gupakira
    • Gukoresha kwibuka cyane
  6. Ni irihe terambere WGAN-GP yazanye kuri WGAN y'umwimerere?
    • Imashini nini
    • Igihano gisanzwe gisimbuza gukuramo ibiro
    • Imibare mishya
    • Gushimangira ibihembo byo kwiga
  7. Why can Wasserstein loss provide a useful signal when real and generated distributions barely overlap?
    • It becomes a binary accuracy score
    • It no longer depends on the critic
    • The distance can vary smoothly as the distributions move, providing gradients before they overlap
    • It requires the generator to match every sample exactly
  8. What does the WGAN-GP gradient penalty encourage the critic’s gradient norm to do on sampled interpolations?
    • Become exactly zero everywhere
    • Grow without bound
    • Match the generator’s parameter count
    • Stay near one