运行和验证

ScaNN编译安装完成后获取到测试数据集进行ScaNN的运行和验证。

1. 进入ScaNN功能验证规划路径。

   1	cd /path/to/scann_test

2. 下载测试数据集。

   1	wget http://ann-benchmarks.com/glove-100-angular.hdf5 --no-check-certificate

3. 执行以下命令创建并编写ScaNN测试脚本scann_test.py。
   1. 创建“scann_test.py”文件。

      1	vi scann_test.py

   2. 按“i”进入编辑模式，编写“scann_test.py”文件，添加如下部分。

      import numpy as np
      import h5py
      import time
      
      import scann
      
      def compute_recall(neighbors, true_neighbors):
          total = 0
          for gt_row, row in zip(true_neighbors, neighbors):
              total += np.intersect1d(gt_row, row).shape[0]
          return total / true_neighbors.size
      
      def main():
          print("Load dataset: glove-100-angular.hdf5")
          glove_h5py = h5py.File("glove-100-angular.hdf5", "r")
      
          print("Dataset keys:", list(glove_h5py.keys()))
          dataset = glove_h5py['train']
          queries = glove_h5py['test']
          print("Train size: ", dataset.shape)
          print("Queries size:", queries.shape)
      
          print("\nCreate ScaNN searcher")
          start = time.time()
          normalized_dataset = dataset / np.linalg.norm(dataset, axis=1)[:, np.newaxis]
          searcher = scann.scann_ops_pybind.builder(normalized_dataset, 10, "dot_product").tree(
              num_leaves=2000, num_leaves_to_search=100, training_sample_size=250000).score_ah(
              2, anisotropic_quantization_threshold=0.2).reorder(100).build()
          end = time.time()
          print("Time (s):", end - start)
      
          print("\n1.Batched-query: queries")
          start = time.time()
          neighbors, distances = searcher.search_batched(queries)
          end = time.time()
          print("Recall:", compute_recall(neighbors, glove_h5py['neighbors'][:, :10]))
          print("Time (s):", end - start)
      
          print("\n2.Single-query: queries[0]")
          start = time.time()
          neighbors, distances = searcher.search(queries[0], final_num_neighbors=5)
          end = time.time()
      
          print("neighbors:", neighbors)
          print("distances:", distances)
          print("Time (ms):", 1000*(end - start))
      
      if __name__ == "__main__":
          main()
      

   3. 按“Esc”键，输入:wq!，按“Enter”保存并退出编辑。
4. 运行测试。

   1	python3 scann_test.py

   

   回显信息显示，测试程序先加载数据集glove-100-angular（该数据集为100维，训练集规模约100万条，查询数据集为10000条），之后创建一个ScaNN的searcher搜索器，最后采用2种模式查询数据，如下：

   1. Batched模式：批量查询全部查询数据集，在该参数配置下，召回率Recall为0.89965。
   2. Single模式：查询数据集中索引为0的数据，结果返回与其近似最近的5个邻居的位置（neighbors）和对应距离信息（distances）。

   若测试程序运行无报错，Batched模式下的召回率与上图回显信息相近，Single模式查询的数据信息与上图回显信息一致，则代表ScaNN功能正常。