search

zoubohao / DenoisingDiffusionProbabilityModel-ddpm-

This may be the simplest implement of DDPM. You can directly run Main.py to train the UNet on CIFAR-10 dataset and see the amazing process of denoising.
MIT License
1.48k stars 156 forks source link

what do channel and channel_mult mean? #4

Open junbopeng opened 1 year ago

junbopeng commented 1 year ago

what do channel and channel_mult in the Main.py mean?

AroundKiki commented 1 year ago

I used to have this question as well, although it was some time ago. These two parameters control U-Net architecture. The first one is 'channel', which corresponds to the output channel of the first layer in U-net. Secondly, in U-net architecture, the channel of each layer gradually increases by 'channel_mult'. This translates to a sequence of channels that goes like [channel ch_mult[0], channel ch_mult[1], ...]. These two parameters should be tuned by your task.