How to calculate NLL (which is mentioned in the paper) ?

[Josh00-Lu]

Are there any explanations?

[sndnyang]

I found one related work https://github.com/baofff/Extended-Analytic-DPM the api of nll is https://github.com/baofff/Extended-Analytic-DPM/blob/main/interface/evaluators/dtdpm_evaluator.py#L78 Then it calls the function in https://github.com/baofff/Extended-Analytic-DPM/blob/main/core/diffusion/likelihood.py

But I'm not sure if its evaluation is the same as DDPM/iDDPM

[sndnyang]

Another one is https://github.com/openai/guided-diffusion/blob/main/guided_diffusion/gaussian_diffusion.py#L709

[sndnyang]

Another one is https://github.com/openai/guided-diffusion/blob/main/guided_diffusion/gaussian_diffusion.py#L709

It starts from guided_diffusion/scripts/image_nll.py

This is easy to follow and the result makes sense (close to the table reported)

[Josh00-Lu]

Another one is https://github.com/openai/guided-diffusion/blob/main/guided_diffusion/gaussian_diffusion.py#L709

It starts from guided_diffusion/scripts/image_nll.py

This is easy to follow and the result makes sense (close to the table reported)

Thanks a lot! It's a pity that the authors didn't explain in detail about how to implement the "NLL metric" in their paper. It's easy to get confused when directly follow others' PyTorch code.

As far as I know, to calculate NLL, we need to calculate the $\sumi log(P\theta (x{real}^{i}))$, while I don't know how to get $P\theta(x)$ (e.g. the PDF of the distribution of our model)? I'm new to this field.

[Josh00-Lu]

What if there is a "maths explanation" will be better! Whatever, thanks a lot!

[sndnyang]

What if there is a "maths explanation" will be better! Whatever, thanks a lot!

A math explanation is in DDPM http://arxiv.org/abs/2006.11239, section 3.3 Data scaling, reverse process decoder, and

[YuanYuan98]

Another one is https://github.com/openai/guided-diffusion/blob/main/guided_diffusion/gaussian_diffusion.py#L709

It starts from guided_diffusion/scripts/image_nll.py

This is easy to follow and the result makes sense (close to the table reported)

Thanks for the useful pointer. The function _vb_terms_bpd calculates the NLL only for one step. If the NLL of one sample is required, should I sum up the results of _vb_terms_bpd for all diffusion steps (t=0-T)?

[sndnyang]

Another one is https://github.com/openai/guided-diffusion/blob/main/guided_diffusion/gaussian_diffusion.py#L709

It starts from guided_diffusion/scripts/image_nll.py This is easy to follow and the result makes sense (close to the table reported)

Thanks for the useful pointer. The function _vb_terms_bpd calculates the NLL only for one step. If the NLL of one sample is required, should I sum up the results of _vb_terms_bpd for all diffusion steps (t=0-T)?

Maybe you can check my wrapper code based on OpenAI's implementation https://github.com/sndnyang/iDDPM

[luccachiang]

Another one is https://github.com/openai/guided-diffusion/blob/main/guided_diffusion/gaussian_diffusion.py#L709

It starts from guided_diffusion/scripts/image_nll.py This is easy to follow and the result makes sense (close to the table reported)

Thanks for the useful pointer. The function _vb_terms_bpd calculates the NLL only for one step. If the NLL of one sample is required, should I sum up the results of _vb_terms_bpd for all diffusion steps (t=0-T)?

Maybe you can check my wrapper code based on OpenAI's implementation https://github.com/sndnyang/iDDPM

Thanks for your comprehensive guide! May I ask what is the relationship between nll and bpd? Do you have any explanations? Thx in advance.