search

Seanseattle / SMIS

Semantically Multi-modal Image Synthesis(CVPR 2020)
Other
322 stars 48 forks source link

Some problems about ADE20K training #13

Open VicZlq opened 3 years ago

VicZlq commented 3 years ago

Hello! Thank you for your wonderful project! When I trained ADE20k data set, some problems appeared. What is the value of-use _ vae10 in ade20k.sh? I set 10 to start training, but the speed is very slow. What equipment do you train on? About how long did you train? Thank you! Looking forward to your reply!

jessemelpolio commented 3 years ago

Training on such a dataset with so many classes actually requires a huge budget of computation. It is because setting so many groups for a group conv would hugely decrease its speed. A very simple fix is to reduce the group number to a reasonable value by grouping some semantic classes together. I remember the training time for ADE20K was around a week on a machine with 8 * 32GB Tesla V100 GPUs.