phquang/continual-normalization

continual-normalization's Issues

_CN equivalent implementation

Hi,

I'm trying to use this implementation in my architecture (Class Incremental Learning), and I have a doubt regarding the CN layer proposed.

Since GN and BN are not sharing weights, can we split these into two layers? CN

Proposed :

class _CN(nn.Module):
    def __init__(
        self,
        num_features: int,
        momentum: float = 0.01,
        affine: bool = True,
        n_groups=32,
        name="2d",
        eps: float = 1e-3,
    ):
        super().__init__()
        self.gn = nn.GroupNorm(
            num_channels=num_features, num_groups=n_groups, eps=eps, affine=affine
        )
        self.bn = {"1d": BatchNorm1d, "2d": BatchNorm2d, "3d": BatchNorm3d}[name](
            num_features, eps, momentum, affine
        )

    def forward(self, X):
        return self.bn(self.gn(X))

class CN32BN2(_CN):
    def __init__(
        self,
        num_features: int,
        momentum: float = 0.01,
        affine: bool = True,
        eps: float = 1e-3,
    ):
        super().__init__(
            num_features, momentum=momentum, affine=affine, n_groups=32, name="2d", eps=eps
        )

Reproduction of Experiments

Hello,
We are trying to reproduce some of experiments in the Continual Normalization: Rethinking Batch Normalization For Online Continual Learning paper. We will have some questions about the paper and repository.

Are there any exact list of hyperparameters of the experiments? We couldn't reproduce published results. In addition, how can we achieve datasets and model settings for Long-Tailed Online Continual Learning experiments which made in Section 5.3. NUS-WIDEseq and COCOseq benchmarks , model and hyperparameters are not available in source code and mammoth framework.

Best,

Atakan Site, Mehmet Selahaddin Şentop

When will your code be released？

Issue about the provided example

Hi Thanks for the great work, It is really simple and elegant. But I have a question about the provided example. In your example codes, what is the variable nl_fn in line 70 (last line of function replace_bn)?