patching supervised loss weight because it is not passed to NeuralAdmixture class

[rochevin]

Hello again,

Thanks again for answering to my issue.

I noticed that the --supervised_loss_weight parameter is available to modify the supervised loss weight. However, it doesn’t seem to propagate correctly into the NeuralAdmixture class, the code always uses the default value of 100.

I applied a small patch to fix this and confirmed that the parameter now behaves as expected on my dataset.

My commit is just a suggestion, please feel free to use or adapt it if you find it helpful.

Examples

In my dataset, when I'm using the pip installed version of neural-admixture (1.6.7) :

for e in 100 200 500 1000 2000 5000 10000 20000 50000 100000; do
    neural-admixture train --k 11 --data_path ../results/20250801/genotypes/20250801_popref.bed --threads 20 --save_dir . --num_gpus 1 --name neural_20250801_svp_train_500_${e} --pops_path ../results/20250801/genotypes/20250801_popref.neuralpop --supervised_loss_weight ${e} --epochs 500
done

Each facet on the second plot I've show the Q matrix for the training dataset when varying the --supervised_loss_weight parameter but as you can see nothing change. On the top plot you can see the result with classic Admixture as a comparison.

And with the patch I did:

Best regards,
Vincent Rocher