zjulearning / ssg Goto Github PK

Hi, I want to run SSG on Deep100m dataset.
Would you kindly send me the link to the dataset, please?
Also, can you share Faiss, efanna refine, SSG/NSG config for Deep100m?
Thanks,

root@f04dcf1c3705:/opt/SSG/build/tests# ./test_ssg_index /opt/SSG/data/datasets/sift/sift_base.fvecs /opt/SSG/data/datasets/knn/sift_200nn.knng 200 200 12 10 100
Using Seed 100
Data Path: /opt/SSG/data/datasets/sift/sift_base.fvecs
L = 200, R = 200, Angle = 12
KNNG = /opt/SSG/data/datasets/knn/sift_200nn.knng
Output SSG Path: 10
terminate called after throwing an instance of 'std::bad_alloc'
what(): std::bad_alloc
Aborted

请问这个问题您是如何解决的
感谢！

Hi,

I am interested in your work and want to use ssg graph to my own implemented search algorithm.

I want to know this graph is stored? What is it data layout or looks like?

Can I convert it into a general graph format and dense matrix/ adjacent list/ CSR something like that?

请教为什么ssg这么计算夹角？我理解计算cone圆锥的夹角不是这么计算吧。

float djk = distance_->compare(data_ + dimension_ * (size_t)result[t].id,
data_ + dimension_ * (size_t)p.id,
(unsigned)dimension_);
float cos_ij = (p.distance + result[t].distance - djk) / 2 /
sqrt(p.distance * result[t].distance);
希望能得到指教。

您好，不是很熟悉C++代码，请问可以给出glove转化为二进制文件的工具代码么，感谢~

我是在sift1M上测试的，看论文，两个方法应该效果差不多的，但是我自己没测出来:(
1）NSSG的构建时间是32.9728+35.99(s)；最关注的搜索时间是search_L=100, k=10,时间是17.3324s，而NSG可以3.29968s跑出来,显然是NSG的QPS高了；
2）在我设置多个参数，绘制 QPS vs Recall@10 曲线时，SSG也确实没有outperformNSG；
请问您能给我一些可能的原因嘛~

From line 29 of CMakeLists.txt:

# Source codes
add_subdirectory(src)
add_subdirectory(tests)

# Compile flags
add_definitions(-std=c++11 -O3 -lboost -ltcmalloc_minimal -march=native -Wall -DINFO)

The correct order should be:

# Compile flags
add_definitions(-std=c++11 -O3 -lboost -ltcmalloc_minimal -march=native -Wall -DINFO)

# Source codes
add_subdirectory(src)
add_subdirectory(tests)

The original order deactivates the compile parameters in fact, which disables the O3 and AVX optimization.

Hello, I'm interested in your work and I'm trying to reproduce your result.

I want to reproduce the result for DEEP100M dataset on single core.
However, test_nn_descent is not working because of segmentation fault.
Can you let me know how did you build kNN graphs for this dataset using Faiss as mentioned in your paper?
Also, can you share both kNN graph and ssg index?

Thanks