Joseph

Btw, as a side note, I found that for more difficult tasks(in my case I'm classifying the output of a mixed speech separator), SincNet is better trained when given, instead...

sounds good, will take a look asap. One question: I only see this guidance used in magic123 and experimental/dreamfusion-sd. Isn't this going to make the changes I propose harder to...

Sounds good. All my code has been pushed. Could you go through the commit and approve it please?

Could we have a quick call? My wechat is 17717613901, would like to clarify a few implementation details On Mon, Dec 18, 2023 at 20:10 Ruizhi Shao ***@***.***> wrote: >...

> conda install -y libffi==3.3 This fixed my problem

In other words, according to this repo's implementation the main SDS loss is not being scaled by the gradient scaler, but other losses are being scaled by the gradient scaler.

confirmed the issue by using the original SDS loss implementation and fixing the loss scaling by setting "self.scaler.update(1.0)". Compared to before when the scaling was not set, the training result...

[image: df_ep0100_0005_rgb.png] [image: df_ep0100_0005_rgb.png] for comparison, first image is using the erroneous implementation of gradient scaling, the second one is setting gradient scaling to 1 thereby disabling it. But this...

 

sudo apt-get install python3-pcl pcl-tools