Discrepancy between paper's loss description and code: MSE on unnormalized embeddings

[gdewael]

Hi!

The preprint describes the mapping loss as MSE on normalized embeddings, which would be equivalent to a cosine loss (2 - 2·cos(y_true, y_pred)). However, in the released code, the normalizations appear to be commented out:

SpecBridge/specbridge/models/mapper.py

Lines 151 to 157 in 6c6fadf

	# z_s_n = F.normalize(z_s, dim=-1)
	# z_m_n = F.normalize(z_m, dim=-1)
	# mu_n = F.normalize(mu_s, dim=-1)
	z_s_n = z_s
	z_m_n = z_m
	mu_n = mu_s

Similarly, the molecule embedding projection + normalization is commented out at line 135

SpecBridge/specbridge/models/mapper.py

Line 135 in 6c6fadf

# z_m = F.normalize(self.chem_proj(z_m), dim=-1)

This means the MSE loss operates on unnormalized embeddings, which is not equivalent to cosine loss.

SpecBridge/specbridge/models/mapper.py

Line 171 in 6c6fadf

L_map = F.mse_loss(mu_n, z_m_n, reduction='mean') # * mu_n.size(1)

Could you clarify:

1. If I'm overlooking something?
2. If the released checkpoints correspond to training with normalized or unnormalized MSE?
3. If the paper description reflects the intended final configuration, or were the normalizations intentionally disabled?

Thanks!