Question about reproducing the best result and pretrained weights

[jikerWRN]

Hi authors,

First of all, thank you for the great work on SmellNet. The dataset and benchmark are very valuable for sensor-based machine
olfaction, and the combination of low-cost MOX sensor signals with GC-MS supervision is especially interesting.

I tried to reproduce the best Transformer / ScentFormer result reported in the paper for the SmellNet-Base classification
task. According to the paper, the best setting uses cross-modal GC-MS supervision with:

• model: Transformer / ScentFormer
• window size: 100
• temporal differencing: 25
• epochs: 90
• batch size: 32
• learning rates searched: 0.0003, 0.001, 0.003

I ran the following script to compare the three learning rates with two random seeds:

cd models
for lr in 0.0003 0.001 0.003; do
    python run.py \
      --train-dir ../base_data/training \
      --test-dir ../base_data/testing \
      --real-test-dir ../base_data/testing \
      --gcms-csv ../gcms_processed/gcms_food_vectors.csv \
      --models transformer \
      --contrastive on \
      --window-sizes 100 \
      --gradients 25 \
      --epochs 90 \
      --batch-size 32 \
      --lr "$lr" \
      --seed "$seed" \
      --run-name-prefix "transformer_lr${lr}_seed${seed}"
  done
done

The best Acc@1 I obtained was around 59.1%, while the paper reports 63.3% for ScentFormer with GC-MS supervision. I was
wondering if you could help clarify what might explain this gap.

Finally, would it be possible to release the pretrained model weights for the reported best ScentFormer / Transformer model and other models?
Having the weights available would be very helpful for reproducibility and would be important for future research building
on this benchmark.

Thank you again for releasing this work!