Open KakaruHayate opened 3 weeks ago
@KakaruHayate Hi, thanks for your interest in our project. This discriminator is composed of multiple sub-discriminators. Since it applies multi-window and multi-band analysis simultaneously, the number of the total sub-discriminators is (number of bands) (number of windows) and in the paper, this number is configured as 35=15. Besides, each sub-discriminator has an accompanied PatchGAN discriminator. Therefore, there are a total of 30 sub-discriminators in the model.
You can reduce the number of windows to improve the speed.
@KakaruHayate Hi, thanks for your interest in our project. This discriminator is composed of multiple sub-discriminators. Since it applies multi-window and multi-band analysis simultaneously, the number of the total sub-discriminators is (number of bands) (number of windows) and in the paper, this number is configured as 35=15. Besides, each sub-discriminator has an accompanied PatchGAN discriminator. Therefore, there are a total of 30 sub-discriminators in the model.
You can reduce the number of windows to improve the speed.
Thank you for your reply. I noticed that time_lengths and freq_lengths were not actually used Because they are not necessary?
Please check the MultiWindowDiscriminator class. The time_lengths
is used to initialize multiple DiscriminatorFactory
instances. These instances are bundled to the self.conv_layers
attribute.
Hello, thank you very much for your work
I am trying to transplant the MB-discriminator to other projects, but the discriminator is very time-consuming during training.
I used the above method for testing, and the discriminator takes 2 seconds each time. Does this comply with the design? Or are there still some parts that have not been updated? I am testing on RTX A4000 (16g VRAM).