Umbala we-Left ABSTRACT Ukuhlobisa 1 INTRODUCTION 1 Ukulungiswa 2 INTERACTIVE WORLD SIMULATION 2 I-Interactive World Simulation 3 GAMENGEN 3 Imidlalo I-Data Collection nge-Agent Play 3.2 Ukwakhiwa kwe-Generative Diffusion Model 4 EXPERIMENTAL SETUP 4 Isakhiwo se-experiment 4.1 Uqeqesho we-Agent 4.2 Generative Model Ukuqeqeshwa 5 RESULTS 5 Iziphumo 5.1 Umgangatho we-simulation 5.2 Ukukhishwa 6 RELATED WORK 6 Ukusebenza okuqukethwe 7 DISCUSSION, ACKNOWLEDGEMENTS AND REFERENCES 7 Ubuchwepheshe, Imibuzo futhi Imibuzo 4 EXPERIMENTAL SETUP 4 Isakhiwo se-experiment 4.1 AGENT TRAINING I-agent model ifakwe usebenzisa i-PPO (Schulman et al., 2017), nge-CNN efanelekayo njengenethiwekhi ye-feature, ngokulandelana ne-Mnih et al. (2015). It is trained on CPU usebenzisa i-Infrastructure Stable Baselines 3 (Raffin et al., 2021). I-agent ifakwe nge-versions ezincinane ze-frame images kanye ne-in-game map, ngamunye ku-resolution 160x120. I-agent futhi inezinto ezingu-32 ezidlulayo. Inethiwekhi ye-feature ivimbela ukubonisa ubukhulu we-512 ngalinye imifanekiso. Umbhali we-PPO no-critic yi-2 layer MLP heads phezulu kwe-concatenation ye-outputs ye-image feature network ne- 4.2 GENERATIVE MODEL TRAINING Thola zonke amamodeli zokuxhumana kusuka ku-Data Pre-Trained Checkpoint of Stable Diffusion 1.4, ukunciphisa zonke imingcele ye-U-Net. Thola usebenzisa usayizi ye-128 futhi isivinini sokufundisa okuqhubekayo se-2e-5, ne-Adaptor Optimizer ngaphandle kokubilisa (Shazeer & Stern, 2018) ne-gradient clipping of 1.0. Thola iziphumo zokusebenza kwe-diffusion loss parameterization ku-v-prediction (Salimans & Ho (2022a). I-context frames condition is eyenziwe nge-probability 0.1 ukuze kuvumela i-CFG ngesikhathi sokufunda. Thola usebenzisa 128 TPU-v5e izixhobo nge-data parallelization. Ngaphandle kokubizwa okunye, zonke iziphumo emakh 
 
 
 Umbhali: 
 
 
 
 
 Dani Valevski Yaniv Leviathan Moab Arar Shlomi Fruchter Umbhali: 
 
 
 
 
 U-Dani Valevski Umculo we-Leviathan I-Moab Arar Shlomi Fruchter 
 
 Okuzenzakalelayo iyatholakala ku-archiv ngaphansi kwe-license ye-CC by 4.0 Deed (i-Attribution 4.0 International). Okuzenzakalelayo iyatholakala ku-archiv ngaphansi kwe-license ye-CC by 4.0 Deed (i-Attribution 4.0 International). I-Archive ye-Archive

Part of HackerNoon's growing list of open-source research papers, promoting free access to academic material.

HOMEPAGE: Interoperability.TECH

Interoperability

Lo msindo ukhiqizwa ngolimi lokuqala lwendaba!

Indlela Ukufundisa Ukufundisa kanye nokuhlukaniswa okuqhubekayo zihlanganisa izwe ze-Game Worlds

About Author

IMIBONO

HANG TAGS

LESI SIHLOKO SETHULWE NGAPHAKATHI

Related Stories

Zulu Network: Moving The Bitcoin Economy Forward With a Two-Tiered Bitcoin Layer 2 Architecture

The Billionaires F*cked Up

Zulu Network: Moving The Bitcoin Economy Forward With a Two-Tiered Bitcoin Layer 2 Architecture

The Billionaires F*cked Up

Light-Mode

Classic

Newspaper

Minty

Dark-Mode

Neon Noir

Minty

HN StartUps