Question regarding MOSAIKS Implementation

Issue に示されている RCF クラスから始め、pull request #70 の実装を調査してください。コードのランダム畳み込み処理を、MOSAIK 論文におけるパッチサンプリングの説明と比較してください。差異が意図的なものかどうかを判断してその根拠を文書化するか、必要な実装変更を特定できれば完了です。

Discrepancy between MOSAIK paper and MPC RCF implementation

Hi!

Given the recent open-sourcing of AlphaEarth by DeepMind and its comparison with MOSAIK [1], I was taking a closer look at the implementation from @calebrob6's PR (https://github.com/microsoft/PlanetaryComputerExamples/pull/70).

Expected behavior (from paper)

Based on the MOSAIK paper, the Random Convolution Features (RCF) process works as follows:

1. Given an input image I, features are computed by randomly sampling K patches from across N images from the training set
2. These K patches are then convolved over I to obtain K feature maps
3. The feature maps are averaged over pixels to produce a K-dimensional feature vector X_i

The following image from the paper illustrates this process:

Actual implementation

However, the MPC implementation uses a different approach:

class RCF(nn.Module):
    """A model for extracting Random Convolution Features (RCF) from input imagery."""
    def __init__(self, num_features=16, kernel_size=3, num_input_channels=3):
        super(RCF, self).__init__()
        # We create `num_features / 2` filters so require `num_features` to be divisible by 2
        assert num_features % 2 == 0
        self.conv1 = nn.Conv2d(
            num_input_channels,
            num_features // 2,
            kernel_size=kernel_size,
            stride=1,
            padding=0,
            dilation=1,
            bias=True,
        )
        nn.init.normal_(self.conv1.weight, mean=0.0, std=1.0)
        nn.init.constant_(self.conv1.bias, -1.0)
    
    def forward(self, x):
        x1a = F.relu(self.conv1(x), inplace=True)
        x1b = F.relu(-self.conv1(x), inplace=True)
        x1a = F.adaptive_avg_pool2d(x1a, (1, 1)).squeeze()
        x1b = F.adaptive_avg_pool2d(x1b, (1, 1)).squeeze()
        if len(x1a.shape) == 1:  # case where we passed a single input
            return torch.cat((x1a, x1b), dim=0)
        elif len(x1a.shape) == 2:  # case where we passed a batch of > 1 inputs
            return torch.cat((x1a, x1b), dim=1)

As I understand it, instead of using K kernels extracted from N training images (as described in the paper), this implementation creates K randomly initialized kernels. While randomly initialized kernels can be powerful feature extractors, this approach differs significantly from the paper's methodology.

Am I missing something that explains why this implementation choice was made? Is there a specific reason for deviating from the paper's approach of sampling patches from training images?

Thank you!

References
[1] Rolf, Esther, et al. "A generalizable and accessible approach to machine learning with global satellite imagery." Nature communications 12.1 (2021): 4392.