Why use Conv1d(k=1) as the layer aggregator in the inference demo?

[kmraven]

Hi authors,

Thank you for this great job.
I have a question about the use of nn.Conv1d(in_channels=13, out_channels=1, kernel_size=1) as an "aggregator" for layer-wise hidden states (shape: [13, 768]) in scripts/MERT_demo_inference.py

As far as I understand, this Conv1d is not trained in the demo—so it acts as a linear layer with random initialization. In this case:

• What is the meaning for using a randomly initialized Conv1d here?
• Wouldn't a simple "uniform average" of the 13 layer representations (i.e., summing and dividing by 13) be sufficient?
• Or, is this a placeholder for fine-tuning later?

I’d appreciate any insights. Thank you!