Maswali yanayounganishwa na mwongozo · Kati Kiwango

Maswali ya Miundo ya Uthabiti

Elewa jinsi miundo ya uthabiti hujifunza kizazi cha hatua moja au chache kutoka kwa njia za uenezi na vikwazo vya mipaka.

Njia za mwongozo zinazohusianaMifano ya Uthabiti
Swali 1 ya 8

Ni shida gani kuu na mifano ya uenezi ambayo mifano ya uthabiti hushughulikia?

Hakiki kila swali8 maswali

Majibu na maelezo huonekana unapojibu maswali.

  1. Ni shida gani kuu na mifano ya uenezi ambayo mifano ya uthabiti hushughulikia?
    • Sampuli za uenezi ni polepole kwa sababu zinahitaji hatua nyingi za kurudia
    • Usambazaji hauwezi kutoa picha za rangi
    • Usambazaji hutumia kumbukumbu ndogo sana
    • Usambazaji hufanya kazi kwenye maandishi pekee
  2. Je, ni mali gani ya 'kujisimamia' ambayo mifano hii imefunzwa kukidhi?
    • Kila picha lazima iwe na rangi sawa
    • Kila nukta kwenye ramani za mwelekeo wa kelele-hadi-data hadi sehemu safi sawa
    • Mtindo lazima akubaliane na mtu anayeweka lebo
    • Pato lazima lilingane na kelele ya ingizo
  3. Je, ni hali gani ya mpaka ambayo kitendakazi cha uthabiti kinapaswa kukidhi kwa wakati sufuri?
    • Ni lazima kutoa kelele safi
    • Ni lazima itoe picha nyeusi
    • Ni lazima ifanye kazi kama kitambulisho (rejesha ingizo bila kubadilika)
    • Ni lazima itoe wastani wa seti ya data
  4. Mfano wa uthabiti unawezaje kuundwa kutoka kwa mfano uliopo wa uenezaji?
    • Kwa kufuta nusu ya tabaka zake bila mpangilio
    • Kwa kuibadilisha kuwa mti wa uamuzi
    • Kwa kuifundisha tu kwenye data ya maandishi
    • Kwa kunereka kwa uthabiti wa modeli ya uenezi iliyofundishwa
  5. Ni mbinu gani hudumisha mafunzo ya uthabiti kwa kutoa malengo ya kujifunza?
    • Mtandao unaolengwa umesasishwa kama wastani wa kusonga mbele
    • Inafuta kitendakazi cha upotezaji bila mpangilio
    • Kuongeza kiwango cha kujifunza mara mbili kwa kila hatua
    • Kutumia picha moja tu ya mafunzo
  6. How many sampling steps did the Latent Consistency Models paper report for high-quality 768×768 image generation?
    • Hundreds of steps in every setting
    • Two to four sampling steps
    • Exactly one training epoch
    • No sampling or model evaluation
  7. What are two training routes introduced for consistency models?
    • Only supervised image classification
    • Only adversarial training with a discriminator
    • Consistency distillation from a pretrained diffusion model and standalone consistency training
    • Token prediction and masked language modeling
  8. Which zero-shot image-editing tasks did the consistency-model paper demonstrate?
    • Speech diarization and speaker verification
    • Machine translation and summarization
    • Database query planning and code completion
    • Inpainting, colorization and super-resolution