I have recently been trying to reproduce your work, but I noticed that in this repository, the VAE encoder's gradient is accidentally truncated. This seems to bring a substantial performance boost. However, this implementation differs from Equation (7) in the paper, which does not mention any gradient truncation for the encoder. In other words, the gradient of the second term in the "gluing of z_0" is not taken into account. Have you investigated this phenomenon? Thank you very much!
I have recently been trying to reproduce your work, but I noticed that in this repository, the VAE encoder's gradient is accidentally truncated. This seems to bring a substantial performance boost. However, this implementation differs from Equation (7) in the paper, which does not mention any gradient truncation for the encoder. In other words, the gradient of the second term in the "gluing of z_0" is not taken into account. Have you investigated this phenomenon? Thank you very much!