yang-song / score_flow Goto Github PK

Dear Author,

Many thanks for such amazing works. I am very interested in your code.

Do you have any plan to release the Pytorch Version in the future?

Thanks again.

Dear authors,

I'm using your pretrained checkpoints in the google drive. My config file is the deep vp file. However, when running at the following lines in run_lib.py:

    try:
      state = checkpoints.restore_checkpoint(checkpoint_dir, state, step=ckpt)
    except:
      time.sleep(60)
      try:
        state = checkpoints.restore_checkpoint(checkpoint_dir, state, step=ckpt)
      except:
        time.sleep(120)
        state = checkpoints.restore_checkpoint(checkpoint_dir, state, step=ckpt)

It raise an error of "KeyError: 'Conv_2'".

How can I solve this problem? Thank you!

Dear authors,

I'm using the pretrained model "cifar10_deep_vp_likelihood_iw_flip", and I get this error when restoring the checkpoint:

flax.errors.ScopeParamShapeError: Inconsistent shapes between value and initializer for parameter "scale" in "/ResnetBlockBigGANpp_0/GroupNorm_0": (1, 1, 1, 128), (128,). (https://flax.readthedocs.io/en/latest/flax.errors.html#flax.errors.ScopeParamShapeError)

What should I do? Thank you!

Dear authors,

Is there a clean version of the "requirements" file? The current file seems to have many useless packages, such as "ufw".

Hi,

Thanks for providing the code and an interesting paper.

I have a question regarding deriving the score-matching objective via minimizing the KL divergence (Theorem 2). Can you let me know how using the IBP, the terminal terms got canceled or modified to get this objective (in (i) in the image attached) and what $h_{p}^{\top}(x)$ denotes in the following equation in the equation below, as it's not mentioned in theorem anywhere?

Regards,
Yogesh