VAR icon indicating copy to clipboard operation
VAR copied to clipboard

Training code for VAE

Open kl2004 opened this issue 1 year ago • 22 comments
trafficstars

Thanks for the great work and the released code base! 💯

After checking current training code, I notice that the VAE is loading from the pretrained checkpoint vae_ch160v4096z32.pth.

As mentioned #5, the training code for VAE will be released at https://github.com/FoundationVision/vae-pro. Is there a target date for it? I'm very interested in training a VAE model on my custom dataset.

kl2004 avatar Apr 19 '24 21:04 kl2004

Same request here, @FoundationVision any plan to release the VAE training code?

luohao123 avatar Apr 23 '24 06:04 luohao123

@kl2004 @kl2004 we're actively cleaning the codes now. For a temporary reference you can refer to the VQVAE forward at https://github.com/FoundationVision/VAR/blob/main/models/vqvae.py#L56-L59.

keyu-tian avatar Apr 23 '24 11:04 keyu-tian

thanks,

def forward(self, inp, ret_usages=False):   # -> rec_B3HW, idx_N, loss
        VectorQuantizer2.forward
        h_BChw, usages, vq_loss, mean_entropy_loss = self.quantize(self.quant_conv(self.encoder(inp)), ret_usages=ret_usages)
        return self.decoder(self.post_quant_conv(h_BChw)), usages, vq_loss, mean_entropy_loss

does this runable code?

luohao123 avatar Apr 24 '24 04:04 luohao123

I would like to if the evaluation (e.g. FID LPIPS PSNR, etc.) code of VAE can be released?

RobertLuo1 avatar Apr 28 '24 04:04 RobertLuo1

Still looking forward to it...

eanson023 avatar May 02 '24 04:05 eanson023

Thanks for the great work @FoundationVision , still looking forward to the training code for multi-res VQVAE!

chikunw avatar May 06 '24 11:05 chikunw

Still lokking forward

FanqingM avatar May 14 '24 10:05 FanqingM

Still looking forward!

StarCycle avatar May 15 '24 04:05 StarCycle

Still looking forward!

Junda24 avatar May 15 '24 13:05 Junda24

Still looking forward!

HalvesChen avatar May 27 '24 06:05 HalvesChen

Still looking forward!

SunzeY avatar May 27 '24 08:05 SunzeY

每天都来看一眼开源没👁️ 👁️

z379035389 avatar May 31 '24 06:05 z379035389

泪目啊,啥时候能开源

StarCycle avatar May 31 '24 07:05 StarCycle

Dear all @kl2004 @luohao123 @RobertLuo1 @eanson023 @ArmeriaWang @FanqingM @StarCycle @Junda24 @HalvesChen @SunzeY @z379035389

Many thanks for your patience! Our VAE training and inference codebase and model weights would be released next week. Please stay tuned❤️!

I'll post the link here and in the README.

keyu-tian avatar May 31 '24 14:05 keyu-tian

Still looking forward!

HalvesChen avatar Jun 11 '24 07:06 HalvesChen

Dear all @kl2004 @luohao123 @RobertLuo1 @eanson023 @ArmeriaWang @FanqingM @StarCycle @Junda24 @HalvesChen @SunzeY @z379035389

Many thanks for your patience! Our VAE training and inference codebase and model weights would be released next week. Please stay tuned❤️!

I'll post the link here and in the README.

Still looking forward!

walking-shadow avatar Jun 15 '24 05:06 walking-shadow

Still didn't opensouce?

Try ByteDance's work! They have a better work than this

What is the project link?Thanks!

Junda24 avatar Jun 15 '24 11:06 Junda24

Hi @keyu-tian, is the VAE training code being released at https://github.com/FoundationVision/vaex?

kl2004 avatar Jun 25 '24 18:06 kl2004

Still looking forward!

mounchiliu avatar Jul 21 '24 13:07 mounchiliu

Hi @keyu-tian, are you still planning to release the training code? Looking forward to it!

tomekster avatar Aug 05 '24 08:08 tomekster

No, not planing.

luohao123 avatar Aug 06 '24 08:08 luohao123

Hi Guys,

I am trying to train the VQVAE (with multi-scale VQ) using the LlamaGen VAE training code. However, I found the codebook utilization rate remains low (about 58%). Does anyone have a good idea about this? Thanks!

lxa9867 avatar Sep 04 '24 22:09 lxa9867