You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Dear authors,
Very excellent work! When I read the implementation code, I have found the code is slightly different from pseudocode. Take bfnsolver++1 as an example. Does the x_t in code is the variable "μ_i" in pseudocode? Then does the noise_pred is "\hat{x_i}"? If so, in the pseucode, why is there an equation between \hat{x_i} and epsilon?
Sorry for the late reply and for any confusion caused by the code.
To clarify:
x_t in the code corresponds to the variable "μ_i" in the pseudocode.
noise_pred in the code is not "\hat{x_i}" but rather "\hat{\epsilon}".
You can verify this by substituting the expressions for "\bar{alpha}_t" and "\bar{sigma}_t", along with the "\hat{\epsilon}" terms, into the latest iteration equation within the for loop of the pseudocode. This substitution aligns the code with the intended mathematical formulation.
I apologize again for the misleading code. We will update it promptly to ensure consistency with the pseudocode.
Thanks for your reply! By the way, I have another question. The sampling methods based on SDE seem to lack diversity, especially when applied to molecular generation. Is there any good solution or reference? Thank you very much!
Dear authors,
Very excellent work! When I read the implementation code, I have found the code is slightly different from pseudocode. Take bfnsolver++1 as an example. Does the x_t in code is the variable "μ_i" in pseudocode? Then does the noise_pred is "\hat{x_i}"? If so, in the pseucode, why is there an equation between \hat{x_i} and epsilon?
The text was updated successfully, but these errors were encountered: