KeyError: 'qarv_base'

[SingSongbjtu]

Excuse me! Have you ever met the problem of KeyError: 'qarv_base', either training or evaluating(train-var-rate.py or eval-var-rate.py)? Could you please help solve the error?

[ruth-jh]

Have you solve this problem?I have the same question.

[duanzhiihao]

Hi, I'm trying to reproduce the problem, but the code works fine on my end.
Can you share your system information and Python version?

[SingSongbjtu]

Thank you for your reply! I add import lvae.models.qarv.zoo in eval-var-rate.py to run, it seems work. But after that I haven't added import lvae.models.qarv.zoo in train-var-rate.py, it still works well. My python version is 3.9.0, the attachment below is all my configuration.
I have another question hoping to be solve. In the code'plogv = tnf.softplus(plogv + 2.3) - 2.3 # make logscale > -2.3', softplus already makes plogv non-negative since softplus(x)=log(1+exp(x)), why you add another term 2.3 and then minus that? And why you choose the number to be 2.3, can it be an arbitrary number like 5 or 50?
myQARV.txt

[duanzhiihao]

I have another question hoping to be solve. In the code'plogv = tnf.softplus(plogv + 2.3) - 2.3 # make logscale > -2.3', softplus already makes plogv non-negative since softplus(x)=log(1+exp(x)), why you add another term 2.3 and then minus that? And why you choose the number to be 2.3, can it be an arbitrary number like 5 or 50?

The objective is to make scale (i.e., pv) > 0.1, which approximately matches CompressAI. We don't want the scale to be too small, because that causes unstable straining.
We achieve this by first doing plogv = tnf.softplus(plogv + 2.3) - 2.3 to let plogv > -2.3. Since pv = torch.exp(plogv), pv must > 0.1. Note that exp(-2.3) ≈ 0.1.

A much simpler way to achieve this can be pv = plogv.exp() + 0.1. We will probably switch to this simpler way in a future release.