Frontend: Multi-Scale 1D CNN

The frontend processes each detector’s compressed time-domain waveform independently and converts it into a 2D feature map that the backend ResNet can consume.

The two key classes are ConcatBlockConv5 (a single multi-scale inception block) and ConvBlock (three stages of those blocks with downsampling).

ConcatBlockConv5 — the multi-scale block

Each ConcatBlockConv5 applies five parallel 1D convolutions at different kernel scales and concatenates their outputs together with the identity skip connection:

input ──┬── Conv1d(k)    ──┐
        ├── Conv1d(2k)   ──┤
        ├── Conv1d(k//2) ──┤── cat ── Conv1d(1)  ── output
        ├── Conv1d(k//4) ──┤
        ├── Conv1d(4k)   ──┤
        └──────────────────┘  (identity)

All convolutions use padding='same' so that every branch produces the same time length as the input. The five branches capture features at very different time scales: k//4 resolves fast oscillations near merger, while 4k integrates over long inspiral trends. The final 1×1 pointwise convolution fuses the concatenated channels.

Each conv–BN–SiLU sequence uses Conv1dSame so no manual padding calculations are needed.

ConvBlock — the per-detector frontend

ConvBlock stacks three stages of paired ConcatBlockConv5 blocks with spatial downsampling between stages:

input (1, T)
    │
Stage 1: ConcatBlock(k) + ConcatBlock(k//2+1) → MaxPool1d(8)
    │
Stage 2: ConcatBlock(k//2+1) + ConcatBlock(k//4+1) → MaxPool1d(4)
    │
Stage 3: ConcatBlock(k//4+1) + ConcatBlock(k//4+1)
    │
unsqueeze → (1, C, T_down)  — 2D feature map

The unsqueeze at the end adds a height dimension so the output is a 2D image (batch, 1, channels, time_compressed) suitable for the 2D ResNet backend.

One ConvBlock is instantiated per detector, and the outputs are concatenated along the channel dimension before being passed to the backend:

H1 frontend output  (B, 1, C, T_down)
                                        → cat → (B, n_det, C, T_down)
L1 frontend output  (B, 1, C, T_down)

(a) A single multiscale residual block. The input \(x^i\) is analysed simultaneously by \(N_s\) parallel 1D convolutions with kernel sizes \(sk^i\), where \(s\) is a set of constant prefactors applied to the base kernel size \(k^i\); outputs are element-wise summed (\(\oplus\)) with a skip connection. (b) The full frontend architecture: three stages of paired multiscale blocks with progressive spatial downsampling, one ConvBlock per detector.

Parameters

Parameter	Default	Effect
filters_start	32	Base filter count. Stage 1 outputs 32 channels, Stage 2 64, Stage 3 128.
kernel_start	64	Base kernel size k. Larger kernels integrate over more time at the cost of compute and receptive field granularity.
in_channels	1	Input channels per detector (always 1 for the compressed waveform).

Weight initialisation

_initialize_frontend_weights() applies Kaiming-normal initialisation to all Conv1d layers, ones/zeros to all BatchNorm1d layers, and small-normal initialisation to all Linear layers. It is called automatically on construction.