SVG Paths and Commands

[vjagannath786]

One curious after reading the paper, how did you handle Paths and commands. As one command like C might have more representation than other or M can have more representation. So, how did you handle class imbalance.

Also, could you also publish the CE loss for VP-VAE objective.

Thanks.

[vjagannath786]

Looks like the entire obective is changed in the new version of the paper.

[ximinng]

Hi,

Thanks for your follow-up and great questions.

Both MSE and CE losses can be used for the VP-VAE objective. In fact, the CE implementation I used was inspired by the DeepSVG paper. The reason I switched to MSE in the current version is due to a rather unusual CVPR reviewer who insisted that CE doesn't work in this context.

To avoid unnecessary controversy during the review process, I temporarily changed it. However, I’ll make sure to clarify in the final version of the paper that both CE and MSE are viable options.

Best regards,
Ximing

[vjagannath786]

Hi @ximinng,

Thanks for replying, but still I am not fully convinced on the VP-VAE objective. As during the preparation it was said that element and command used embedding layer and parameters are normalized. So, in that case decoder should have separate head for treating element and command as categorical and separate head for numerical parameters.

And, objective becomes a mix of CE loss for element , command and MSE for parameters isn't it ?

Also, possible to post the loss curves of VP-VAE and VP-DIT.