Ci bi mu melGis bi ci topp
Taggat ci waxtu test
Fondamentaal yi
GUIDE teknik
Yokkateg waxtu test
Test-time augmentation (TTA) dafay doxal ab model buñu tàggat ci yenn versioŋ yuñ soppi ci benn duggal bi ba noppi di xayma liñu séentu.
Déglul tegtal biiIA-nettali · ~2 min
Ci xët wii2 simili jàng
Résumé
It is a simple, training-free trick that often squeezes out a few extra points of accuracy and makes predictions more robust.
Plongeur bu xóot
Test-time augmentation dafay jël benn dugal, sos ay kopi yu bari yuñ soppi (flips, crops, rotations, coppite melo, wala versions scaled), dafay daw bu nekk ci benn model fixe, ba noppi boole génne yi - lu bari ci probabilite wala logits. Intuition bi: yokk bu nekk dafay wane model bi ci anam wu wuute tuuti, te njuumte yi ci gis-gis yi dañuy fomm suñu leen boole, melni ensemble bu ndaw buñ tabax ci benn reso. Li gëna am solo mooy TTA soxlawul ñu tàggataat ko, soxlawul yeneen etiket; ci inference bi moo gëna njëg ndax model bi dafay daw N yoon ci échantillon bu nekk. Ñu ngi ko gëna siiw ci wàllu gis-gis ci ordinatër (espesialeman ci joŋante Kaggle ak nataali pajum) waaye dafay feeñ itam ci audio ak bind. Augmentations yi war nañu baña yàq etiket bi - flipping rayon X ci dënn bi baaxna, waaye flipping benn sifre '6' ci '9' baaxul.
Doxal ci àdduna dëgg
Liggéeyu xayma yi ci kaw flip horizontal ak dagg yu bari ci benn nataal ngir gëna dëgëral xaaj ImageNet ci inference.
Delloosi rotation / flips ak mask moyenne ci segmentation nataal medikaal (lu melni, tumër wala cër) ngir gëna dëgër.
Kaggle concurrents jëfandikoo fukki-crop wala multi-scale TTA ngir am benn wàll ci pursaa ci tablo njiit yi te duñu tàggataat.
Dawal wax wala audio classifiers ci kaw clips yu toxal tuuti waxtu wala ton perturbed ak pooling génne ngir etiket yu gëna stabil.
Ëlëgu yokk waxtu test
Gëstu mingi jubal ci TTA buñ jàng te mëna ànd, fu benn politik bu ndaw di tànn augmentations yiy jàppale ci bépp dugal bu amul benn ensemble. Seetug 'Gredy' ak TTA-politik bu wuute, boole ci moyenne bu wóorul bu gëna wóolu xalaat yu wóolu, ay barab yu am solo lañu. Xaarandil TTA ngir jaxasoo ak tàggat yaram ci diiru test ak mën ànd ak sa bopp, bàyyi model yiñ dugal ñu mëna ànd ak coppite séddale ci saa si, boole ci baña soppi moomeel gi.
Teg ko ci jëf
Lu tax mu am solo
- Njëgg ak budget. Dogal yi architecture di jël dañuy indi njariñ ak njëgu liggéey bi ay at ci ginaaw.
- dogal yu gëna leer. Njàngalem xarala yi dafay jàppale ekip yi ñu tànn li gën, te baña yam ci li gëna bees daal.
- Xool kalite. Tanneef yu gëna baax ci wàllu ingeñër dina wàññi jafe-jafe yi ci wàllu wóor ci liggéey bi.
Moytul
- Optimize benn benchmark mën na nëbb ñakk kattan yu gëna yaatu ci sistem bi.
- Njëg li ñuy fay ci infrastructure yi ak ci toppatoo dañuy faral di suufeel.
- Bu sistem yi di gëna xawa jafee xam, jafe-jafe yi am ci wàllu kaaraange ak seetlu mën nañu gëna bari.
Jéego yi ci topp
- Mandargal latency, kalite, ak njëg yi laata ngay jëfandikoo.
- Benchmark ci biir sargal ak done yu dëggu.
- Jumtukaay bi di saytu njuumte yi, derive bi ak njeextalu jëfandikukat bi.
- Waajal rollback ak yooni tontu ci jafe-jafe yi laata ngay eskale.
Weyal di banneexu
Bulletin bu amul fayda
Jàngal xibaar AI bu bes
Ñatti jaar-jaari IA yuñ firndeel bis bu nekk ci ayu-bis bi, ñu bind leen ci Àngle bu leer. Gratis ba fàww, amul yëgle.
Benn email bis bu nekk ci ayu-bis bi. Benn klik nga dindi abonemaa bi. Dunu musa jaay wala séddoo sa adres.
Nattal sa bopp
Take the Test-Time Augmentation quiz
Feedback ci saasi ci bépp tontu, ak sertifikaa buñ mëna séddoo ak ID buñ mëna saytu soo jàllee ci njàng mi.
Jàppale njàngum IA te doo fay. AI Understanding mbootaay la buy def te Yàlla tax — amul yëgle, amul paywall, musul am. Defal ab sarax
Laaj yi ñuy faral di laaj
What is Test-Time Augmentation?
Test-time augmentation (TTA) dafay doxal ab model buñu tàggat ci yenn versioŋ yuñ soppi ci benn duggal bi ba noppi di xayma liñu séentu. Kaf gu yomb la, te soxlawul nga tàggat sa bopp, te dafay faral di dindi yenn poñ yu gëna dëggu, ba noppi gëna dëgëral wax luy waaja am.
At inference time, what does test-time augmentation do with a single input?
TTA dafay dundal ay xeetu yokk yu bari ci benn dugal jaaraleko ci benn xeetu tàggat ba noppi boole li ñu séentu, te amul benn tàggataat bu ci bokk.
When can pooling predictions from transformed views reduce prediction variance?
The ensemble intuition depends on useful, label-preserving views whose errors are not perfectly correlated.
Lan mooy njëgu jëfandikoo TTA?
Ndax dugal bu nekk dafay jaar ci reso bi N yoon, inference dafay gëna seer lu tollu ci N yoon.
Ban tanneef ci yokkute moo mëna gaañ benn klassifikatëru sifar?
Augmentations yi dañu wara tëye etiketu dëgg gi; flipping yenn cifre yi dañuy soppi seen dàntite ba noppi di indi njuumte.
Ci segmentation, ban jéego bu gëna am solo lañu soxla balaa ñuy jël moyenne yi ci TTA yi?
Yokkateg geometrik yi dañuy toxal pixel yi, kon masku génne bu nekk dañu ko wara soppi dellu ci kaadar coordonnée bi njëkk balaa ñuy boole.
Weyal di jàng
Gid yu jëm ci loolu
Tann nañu yeneen njiit ngir topic bii