VSTAR

python3 vstar_generate_models.py --dim <dim> --nlistL1 <nlist1> --subDimL1 <sub_dim1> --nProbeL1 <nprobe1> --nProbeL2 <nprobe2> --segmentNumL3 <segment> --pool <pool_size>

<dim>: feature vector dimension. The default value is 256.

<nlist1>: number of level-1 cluster centroids. The default value is 1024.

<nprobe1>: number of level-1 candidate buckets for each computing on during retrieval. The default value is 72.

<nprobe2>: number of level-2 candidate buckets for each computing during retrieval. The default value is [64, 296].

<sub_dim1>: dimension size after level-1 dimension reduction during retrieval. The default value is 32.

<segment>: number of data segments to be searched for from nprobe2. The default value is [512, 1000, 1504].

<pool_size>: size of the process pool for multi-process scheduling of operators generated in batches. The default value is 16.

--help | -h: help information.

After running this command, you can obtain a group of AI Core and AICPU operator model files used for VSTAR retrieval. You need to modify the parameters in the command.

• dim ∈ {128, 256, 512, 1024}
• nlist1 ∈ {256, 512, 1024}
• sub_dim1 ∈ {32, 64, 128} The value of sub_dim1 must be less than that of dim.
• nprobe1 ∈ (16, nlist1] nprobe1 is a list of the int type, and the value in the list must be an integer multiple of 8.
• nprobe2 ∈ (16, nprobe1 * n] When dim is 1024, n is 16. For other dimensions, n is 32. nprobe2 is a list of the int type, and the value in the list must be an integer multiple of 8.
• segment ∈ (100, 5000] segment is a list of the int type, and the value must be an integer multiple of 8.
• pool_size ∈ [1, 32]. Before running the script, ensure that the maximum number of processes supported by the host is properly set.