Idanwo ti o sopọ mọ itọsọna · Alabọde Ipele

Wasserstein GAN adanwo

Checks Wasserstein distance, critic scores, the Lipschitz condition, weight clipping, gradient penalty and GAN training stability.

Jẹmọ itọsọna awọn ọnaWasserstein GAN
Ibeere 1 ti 8

Metiriki ijinna wo ni WGAN lo lati ṣe afiwe awọn pinpin gidi ati ipilẹṣẹ?

Awotẹlẹ gbogbo ibeere8 ibeere

Awọn idahun ati awọn alaye han bi o ṣe n mu ibeere naa.

  1. Metiriki ijinna wo ni WGAN lo lati ṣe afiwe awọn pinpin gidi ati ipilẹṣẹ?
    • Wasserstein (ayé-mover's) ijinna
    • Kullback-Leibler iyatọ
    • Cosine ibajọra
    • Euclidean ẹbun ijinna
  2. Ni WGAN, nẹtiwọọki ti o jẹ iyasọtọ ni a tunrukọ si kini, ati kilode?
    • 'Adajọ' nitori pe o dibo bẹẹni tabi rara
    • A 'alariwisi' nitori ti o àbájade a lemọlemọfún otito Dimegilio kuku ju iṣeeṣe kan
    • A 'aṣayẹwo' nitori pe o nmu ariwo
    • A 'àlẹmọ' nitori pe o yọ awọn aworan buburu kuro
  3. Kini idi ti alariwisi WGAN gbọdọ ni ihamọ lati jẹ 1-Lipschitz?
    • Lati jẹ ki o ṣe ikẹkọ ni iyara lori awọn GPU
    • Lati din awọn nọmba ti sile
    • Nitoripe Kantorovich-Rubinstein duality ti a lo lati ṣe iṣiro ijinna Wasserstein nilo awọn gradients ti o ni opin
    • Lati fi ipa mu u lati mu awọn nọmba jade
  4. Bawo ni WGAN atilẹba ṣe fi agbara mu idiwọ Lipschitz?
    • Nipa fifi silẹ silẹ
    • Nipa lilo deede ipele nikan
    • Nipa didi monomono
    • Nipa gige awọn iwuwo alariwisi si iwọn kekere kan
  5. Isoro wo ni WGAN dinku pataki ni akawe si awọn GAN fanila?
    • Ipo iṣubu ati isonu ti ko ni alaye
    • Overfitting si awọn igbeyewo ṣeto
    • Ikojọpọ aworan ti o lọra
    • Lilo iranti ti o pọju
  6. Ilọsiwaju wo ni WGAN-GP ṣafihan lori WGAN atilẹba?
    • A o tobi monomono
    • Ijiya onidiwọn ti o rọpo gige gige
    • Titun data
    • Awọn ere ikẹkọ imudara
  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