ThisisBillhe comments

Results 40 comments of


                                            ThisisBillhe

Regarding results

Do you use the imagenet reference batch provided [here](https://github.com/openai/guided-diffusion/tree/main/evaluations)?

quantize my own model to 2bits

You may refer to my other repos [PTQD](https://github.com/ziplab/PTQD) and [torch_quantizer](https://github.com/ThisisBillhe/torch_quantizer) (8-bit only)

Any sample result?

you may refer to our [paper](https://arxiv.org/abs/2310.03270) for visualization results.

I do not train diffusion models from scratch. But you can follow my paper [PTQD](https://github.com/ziplab/PTQD) and [EfficientDM](https://github.com/ThisisBillhe/EfficientDM) to convert a well-trained full-precision diffusion model to a low-bit one.

Will the pre-training code be released in the future?

关于使用 cutlass 实现 per-channel 量化的 int8 conv

Hi，大佬我是不敢当. 要不我们加个微信后续交流？wechat: hyfll2

quantization for performance vs for memory

Hi! You are right, I haven't done any operation fusion yet (both Conv+ReLU and Dequant+Quant). Another reason for slow inference speed is that my dequantization CUDA kernel is slow, which...

Problems on reproducing the experimental results.

any specific config or problem you have met?

Problems on reproducing the experimental results.

1. yes. 2. We use the corresponding FP model as the teacher model for BiViT.

The effect is poor when using the demo for testing.

Hi! Thanks for your interest in our work. I will look into it in a few days.