In the Eq. 24, the $[Hj^{-1}v]=v+[H{j-1}^{-1}v]-\lambda H{\theta{0}}[H_{j-1}^{-1}v]$, and the paper says the the parameter $\lambda$ should be large, such as 10000.
However, in the code, we find that $\lambda==1/scale$
h_estimate = [ v1 + (1-damp)*h_estimate1 - hv1/scale
for v1, h_estimate1, hv1 in zip(v, h_estimate, hv)]
Hi, thanks for the great work!
In the Eq. 24, the $[Hj^{-1}v]=v+[H{j-1}^{-1}v]-\lambda H{\theta{0}}[H_{j-1}^{-1}v]$, and the paper says the the parameter $\lambda$ should be large, such as 10000.
However, in the code, we find that $\lambda==1/scale$
In other word, the scale in the code should be very samll, such as 0.0001. But why this parameter scale should set to be very large in the following issue? https://github.com/wujcan/GIF-torch/issues/1#issuecomment-1566140170