Convolution blocks
The encoder applies four convolution stages with intermittent max pooling. This progressively reduces spatial resolution while expanding channel depth.
Module boundary
`model.py` owns the whole similarity function
The repository keeps all model definition logic inside `model.py`. The `Siamese` class exposes `forward_one` for single-branch feature extraction and `forward` for pairwise scoring.
Encoder path
Shared convolution tower
Weight sharing
Both input images pass through the same `self.conv` and `self.liner` modules, ensuring that similarity is measured in a common embedding space.
Control flow
From `forward_one` to final logit
1. `forward_one`
Each branch image is encoded by the convolution stack, flattened, and projected into a 4096-dimensional embedding with sigmoid activation.
2. Absolute difference
`forward` computes `torch.abs(out1 - out2)`, which is the central similarity comparison primitive in this implementation.
3. Output layer
The difference vector is passed to `self.out`, producing a single logit for binary same-class vs different-class classification.
Current limitations
Implementation tradeoffs in this repo
Single-file ownership
The architecture is compact and easy to read, but the file does not separate encoder, projection, and scoring into independently testable modules.
Minimal output head
The model returns logits directly and leaves probability calibration to the loss function and downstream evaluation code.