Experimental parameters

[lxxabb]

Hello, I encountered some problems while reproducing the experiment and hope to get your answers. The problems are as follows:

1. On the GLOVE100 dataset, when executing the hnsw_training_data_generation.sh script to collect Darth training features, I found that the final accuracy obtained by the parameters you provided, M16, efC=500, efS500, was only 0.8802. This is insufficient to reach 0.95, so I cannot train the model at present. What was the efS parameter you set at that time?

2. In your paper, you mentioned that IPI and MIP are obtained by achieving the required accuracy on the training set (e.g., ...). When accuracy is 0.9, the average distance calculations are used to calculate IPI = distRt/2 and MPI = distRt/10. However, on the GL0VE100 training dataset, with k=50 and accuracy 0.9, efSearch=650. Using efSearch for searching, the calculated average dist is greater than 10,000, but the values ​​you gave in the hnsw_darth_test.sh script are IPI=272 and MPI=54. The difference is quite large. Is my calculation method correct? I would like to ask you for clarification on the specific calculation methods for IPI and MPI.

Thank you for reading. I hope to receive your answer.