Elfina Ashfield

> > solution:关闭mactype > > 还有别的解决办法吗，难到 Clash 和 MacType 就必须二选一吗，具体是哪里冲突了呢？ 可以直接在MacType.ini里面的[UnloadDll]项里面加入 - Clash for Windows.exe - clash-win64.exe - go-tun2socks.exe 来阻止MacType去hook这几个进程。

This config is probably used for inference only. DummyLoss is simply not calculating any loss. Another thing is that kl_weight is ``1e-8``, which makes quantization loss extremely small. I don't...

@borisdayma Because the codebook in the f=8 GumbelVQ model does not contain invalid codes, unlike the IN model. By "invalid codes", I mean: In the IN model's codebook, there are...

You're welcome! @borisdayma I simply treat every code as a 1x1-size patch and forward it through the pretrained decoder.

@borisdayma I haven't been playing around VQ for a while, but hey, we're here. Why not :) Here's the visualization of DALL-E's discrete code (the first 1024 codes of 8192...

> Thanks, @TomoshibiAkira. Where could I find GumbelVQ's model.yaml? I didn't see this config file in this repo. > > Many thanks. It's in the pretrained model zoo. https://heibox.uni-heidelberg.de/d/2e5662443a6b4307b470/?p=%2F&mode=list

> @TomoshibiAkira Don't they both use a codebook where you can use either the codebook index or the corresponding feature vector? @borisdayma I personally don't think so. In the image...

> The DALL-E decoder just uses a simple 1x1 conv2d layer to transform the one-hots into feature vectors (it's a one-to-one mapping), I have opened the decoder up and used...

@fnzhan I didn't conduct the experiment so I can't give any concrete answer. Personally, I'd like to believe that Gumbel can improve the performance without any trade-off since it's basically...

@fnzhan Hmm, that's interesting! This might mean that the actual usage of the codes is very unbalanced no matter the codebook utilization (e.g., the network tends to use several "special"...