lolong

yes, also I got similar results as @ngaloppo reported. openblas + fft16x16 in a i7 machine I0731 21:06:06.510483 12768 caffe.cpp:369] conv2/5x5_s1 forward: 311.694 ms. I0731 21:06:06.510542 12768 caffe.cpp:369] conv3/3x3_s1 forward:...

Ubuntu 20.04, Fastdup version: ```python In [16]: fastdup.__version__ Out[16]: '0.125' ```

with 0.123 still gives this results  i am going to try with the v.126

> hi @lolongcovas I have just released v. 126 for ubuntu 20.04 can you try it out and let me know if this works? You can upgrade using `python3.8 -m...

still seeing the issue. does the fastdup prevent invalid images? ```Wrote total of 999998 components Replacing lower threshold 0.05 with position1899997 top_k.size() 1999996 loc pos: 0.8098 last pos: 00.95 1.9e+06...

the log is huge, i didnt store it as a file. But the `bug` u mention is here: ``` KNN results 0 : 1.00000 14 : 1.00000 814470 : 0.93696...

after removing that bad image, i got this duplicates:  still have the issue. the log: ```python FastDup Software, (C) copyright 2022 Dr. Amir Alush and Dr. Danny Bickson. Going...

now with 0.127 worked. thanks!

hi, sorry again. on the position 187 i found wrong duplicates:  however, 184 and 189 are ok.

I think that there is only one stage. The algorithm trains the cnn model just once. The several stage approach requieres Image alignment as explained in the paper.