How to extract structures and relevant properties after recon, gen and opt tasks?

[chiku-parida]

1. Could you please explain how to extract data from the outputs of recon, gen and opt tasks? @txie-93
2. How will we control the number of generated structures? example: If someone wants to generate 100 structures

[SourinDeyUW]

1. Could you please explain how to extract data from the outputs of recon, gen and opt tasks? @txie-93

   2. How will we control the number of generated structures? example: If someone wants to generate 100 structures

Could you find a solution for it?

[chiku-parida]

I found an alternative way to extract the generated structures. You can load the files as torch tensors. Then you can extract the related tensors like cell parameters, fractional coordinates etc and then using ASE you can create your generated structure using the extracted tensors or the informations. @SourinDeyUW @yqq2022

[SourinDeyUW]

I found an alternative way to extract the generated structures. You can load the files as torch tensors. Then you can extract the related tensors like cell parameters, fractional coordinates etc and then using ASE you can create your generated structure using the extracted tensors or the informations. @SourinDeyUW @yqq2022

Thanks for your response. I have tried.
Could you tell me about how good the reconstructed structures were? Did you have to tune hyperparameters yourself?
I simply ran their code and reconstructed to match with ground truth, resulting cifs look very different. I tried for mp-20 data.

[JHWang1001]

2. How will we control the number of generated structures? example: If someone wants to generate 100 structures

Dear chiku-parida,
Can you solve this problem? When I running cdvae, I only can get 11 structures. However, sometimes I can get 1100 structures. @chiku-parida

[chiku-parida]

@JHWang1001
Hii...
Are you able to solve the issue?
Sorry for late reply. The generation part is working fine for me. I don't understand how it is just generating random samples irrespective of the user input. Can you provide some more details or generation inputs?

[JHWang1001]

@JHWang1001 Hii... Are you able to solve the issue? Sorry for late reply. The generation part is working fine for me. I don't understand how it is just generating random samples irrespective of the user input. Can you provide some more details or generation inputs?

Sorry, I can't solve this issue yet.

When I use task gen, I can find up to 50,000 structures from the tensor! But when I use task opt, this amount seems to depend on the training data. The pervo provided by the author generated 345 structures, while the training set I built myself generated only 11 structures. I still don't seem to understand what is the difference between the gen and opt parameters?

If you want to specify how many structures to output (such as cdvae generates 10,000 structures and you want to produce 50 structures) you can write a loop that extracts just 50 structures. Do you need the code?@chiku-parida

Best wishes
Jiahui Wang

[chiku-parida]

@JHWang1001
The opt task is a part of supervised learning in CDVAE for conditional generation. Somehow I didn't understand the last part you mentioned about the loop that extracts structures. Can you please elaborate? I will look into it again, though I am not using the CDVAE now.