init test framework

CMakeLists.txt

@@ -15,4 +15,4 @@ SHARED
 )
 
 add_executable(hnsw_test src/test.c)
-target_link_libraries(hnsw_test hnswc m)
+target_link_libraries(hnsw_test hnswc)

README.md

@@ -3,9 +3,9 @@
 
 ### 1. Testing Steps
 
-1. Make sure `CMake` is installed:
+1. Make sure `CMake` and other build tools are installed:
    ```shell
-   sudo apt-get install cmake
+   sudo apt-get install cmake build-essentials
    ```
 
 2. Create a `\build` folder inside the `hnswlab` directory.
@@ -27,18 +27,25 @@
 
 6. Run the test program:
    ```shell
-   ./hnsw_tst
+   ./hnsw_test data_file_path data_size query_file_path groundtruth_file_path
    ```
 
-### 2. Interface Description
+   For example:
+   ```shell
+   ./hnsw_test ../dataset/siftsmall/siftsmall_base.fvecs 10000 ../dataset/siftsmall/siftsmall_query.fvecs 100 ../dataset/siftsmall/siftsmall_groundtruth.ivecs
+   ```
+
+### 2. Mission Description
 
-To load the dataset, use the following code snippet in C:
-```c
-// Vector dimension
-int num = 0;
-// Read the dataset
-VecData* vecs = fvecs_read("../dataset/siftsmall_base.fvecs", NULL, &num);
+You need to implement two functions inside hnsw.h and hnsw.c in HNSW way:
+
+```C
+HNSWContext *hnsw_init_context(const char *filename, size_t dim, size_t len);
+void hnsw_approximate_knn(HNSWContext *ctx, VecData *q, int *results, int k);
 ```
 
+We have implemented data loading and provided a simplest KNN algorithm. But our implementation can only handle small batches of data(SIFTSMALL dataset), please implement a new approximate KNN algorithm based on the HNSW algorithm so that it can handle large batches of data(SIFT dataset).
+
+### 3. Data Download
 
----
+Please visit http://corpus-texmex.irisa.fr/ 

inc/hnsw.h

@@ -18,12 +18,12 @@ typedef struct HNSWContext
     HNSWGraph *graph; // graph of HNSW
 } HNSWContext;
 
-// you can define some help functions here
+// you can declare some help functions here, and implement them in 'hnsw.c'
 
 
-// public functions
+// public functions here
 // Please do not modify these function signatures!
 // To simply our program, we do not consider reclaiming memory space here.
-// Please implement these functions according to HNSW algorithm.
+// TODO: Please implement these functions according to HNSW algorithm.
 HNSWContext *hnsw_init_context(const char *filename, size_t dim, size_t len);
 void hnsw_approximate_knn(HNSWContext *ctx, VecData *q, int *results, int k);

inc/utils.h

@@ -19,4 +19,5 @@ float vec_dist(VecData x, VecData y);
 FileContext* init_file_context(const char *filename);
 void read_4bytes(FileContext* ctx, void* dst);
 void read_vec_data(FileContext* ctx, void* dst);
+void read_id_data(FileContext* ctx, void* dst, size_t n);
 void free_file_context(FileContext* ctx);

src/hnsw.c

@@ -1,4 +1,4 @@
-#include <limits.h>
+#include <stdlib.h>
 #include "hnsw.h"
 
 HNSWContext *hnsw_init_context(const char *filename, size_t dim, size_t len)
@@ -35,7 +35,7 @@ void hnsw_approximate_knn(HNSWContext *ctx, VecData *q, int *results, int k)
             if (dist < min_dist)
             {
                 min_dist = dist;
-                idx = i;
+                idx = j;
             }
         }
         if (idx != i)
@@ -47,7 +47,7 @@ void hnsw_approximate_knn(HNSWContext *ctx, VecData *q, int *results, int k)
     }
 
     // copy results
-    for (int i = 0; i < k; i++)
+    for (int i = 0; i < k && i < ctx->len; i++)
     {
         results[i] = ctx->data[i].id;
     }

src/test.c

@@ -3,21 +3,65 @@
 #include "hnsw.h"
 #include "utils.h"
 
+#define K 100
 
-int main(int argc, char *argv[]) {
-    
-    FileContext* f_ctx = init_file_context(argv[1]);
-    VecData data;
-    data.vec = malloc(sizeof(float) * GLOBAL_DIM);
-    for (int i = 0; i < 100; i++)
+float cal_recall_value(int *results, int *trueset, int k)
+{
+    int cnt = 0;
+    for (int i = 0; i < k; i++)
     {
-        read_vec_data(f_ctx, data.vec);
-        for (int j = 0; j < GLOBAL_DIM; j++) 
+        int val = results[i];
+        for (int j = 0; j < k; j++)
         {
-            printf("%f ", data.vec[j]);
+            if (val == trueset[j])
+            {
+                cnt++;
+                break;
+            }
         }
-        putchar('\n');
     }
-    return 0;
+
+    return ((float) cnt) / ((float) k);
 }
 
+int main(int argc, char *argv[])
+{
+    if (argc != 6)
+    {
+        printf("argc: %d\n", argc);
+        printf("./hnsw_test base_file data_size query_file groundtruth_file");
+        exit(1);
+    }
+
+    int data_size = atoi(argv[2]);
+    printf("data size: %d\n", data_size);
+    int query_size = atoi(argv[4]);
+
+    // init query and groundtruth files
+    FileContext *query_file_ctx = init_file_context(argv[3]);
+    FileContext *gt_file_ctx = init_file_context(argv[5]);
+
+    // TODO: report time cost here
+
+    HNSWContext* ctx = hnsw_init_context(argv[1], GLOBAL_DIM, data_size);
+    printf("HNSW context initialized!\n");
+    VecData q_vec;
+    q_vec.vec = (float *) malloc(sizeof(float) * GLOBAL_DIM);
+    int q_results[100];
+    int true_results[100];
+    float total_recall_values = 0.0;
+    for (int i = 0; i < query_size; i++)
+    {
+        read_vec_data(query_file_ctx, q_vec.vec);
+        read_id_data(gt_file_ctx, true_results, K);
+        hnsw_approximate_knn(ctx, &q_vec, q_results, K);
+        total_recall_values += cal_recall_value(q_results, true_results, K);
+    }
+
+    // report recall value
+    printf("Recall value: %.4f\n", total_recall_values / ((float) query_size));
+
+    free_file_context(query_file_ctx);
+    free_file_context(gt_file_ctx);
+    return 0;
+}

src/utils.c

@@ -44,6 +44,13 @@ void read_vec_data(FileContext *ctx, void *dst)
     assert(s == GLOBAL_DIM);
 }
 
+void read_id_data(FileContext *ctx, void *dst, size_t n)
+{
+    read_4bytes(ctx, dst);
+    size_t s = fread(dst, 4L, n, ctx->stream);
+    assert(s == n);
+}
+
 void free_file_context(FileContext* ctx)
 {
     fclose(ctx->stream);