As noticed by @panpiort8, the default USPTO50K-trained RetroKNN weights did not reproduce the USPTO50K accuracy reported in our paper, only getting to ~50% instead of ~55%. It turns out that the weights got mixed up when uploading from our internal storage to figshare. In short, the RetroKNN weights consist of LocalRetro weights and an activation store for retrieval, but the store present in the checkpoint was generated using a different checkpoint of LocalRetro than the one it was uploaded with. After fixing this, the accuracy is back to its expected ~55% range.
As noticed by @panpiort8, the default USPTO50K-trained RetroKNN weights did not reproduce the USPTO50K accuracy reported in our paper, only getting to ~50% instead of ~55%. It turns out that the weights got mixed up when uploading from our internal storage to figshare. In short, the RetroKNN weights consist of LocalRetro weights and an activation store for retrieval, but the store present in the checkpoint was generated using a different checkpoint of LocalRetro than the one it was uploaded with. After fixing this, the accuracy is back to its expected ~55% range.