About lambda

[Hanawh]

Sorry to bother you again. I have another question. I want to know what the value of lambda in loss function depends on. If the picture I use is 112x112, do I need to adjust the lambda in config, or do I need to search for the lambda value

[jbegaint]

Hi, no the lambda is used to balance the two terms in the rate distortion function: the estimated bit-rate and the distortion. The lower the lambda, the more the distortion is penalized over the rate term, which will result in a higher compression rate but a lower reconstruction quality. You may need to search different lambdas to achieve different bit-rates (for example on a logarithmic scale).

[jbegaint]

Closing this for now. Please let us know if you encounter any bug or issue.

[kepengxu]

hello , can you share the diffenent lambda value in quilty 1......8

[jbegaint]

https://interdigitalinc.github.io/CompressAI/zoo.html#training

[Freed-Wu]

Can the diffenent lambda values in quilty 1......8 be provided in compressai/losses/__init__.py? such as

from compressai.losses import q2lambda, RateDistortionLoss
loss_function = RateDistortionLoss(q2lambda(q))

[Freed-Wu]

And can users be allowed to define the output type of RateDistortionLoss? such as

from typing import Literal
class RateDistortionLoss(nn.Module):
    def __init__(self, lmbda: float=1e-2, return_type: Literal["bpp_loss", "mse_loss", "loss", "all"]="all"):
    def forward(self, output, target):
        if self.return_type == "all":
          return out
        else:
          return out[return_type]

[Freed-Wu]

And now RateDistortionLoss only support MSE. Can it support ms-ssim?

[Freed-Wu]

Now compressai model's output is a dict like {"x_hat": torch.Tensor, "likelihoods": XXX}. However many loss function expect only torch.Tensor (i.e, x_hat). Can it provide a decorator to wrap these functions?