Using pytorch-metric-learning and fit_transform() in umap togather cause segementation fault

[ZhuoxuanLiu]

demo code here, test device: macos m2 chip

import numpy as np
from pytorch_metric_learning.utils.accuracy_calculator import AccuracyCalculator
import umap
    
if __name__ == '__main__':
    umapper = umap.UMAP()
    test_embeddings = np.random.normal(size=(100, 64))
    umap_embeddings = umapper.fit_transform(test_embeddings)

env: pytorch-metric-learning == 2.1.0 umap-learn == 0.5.3

[KevinMusgrave]

This sounds like an issue for UMAP, not pytorch-metric-learning. Here are some issues from the umap github repo that might help: https://github.com/lmcinnes/umap/search?q=segmentation+fault&type=issues

[KevinMusgrave]

Do you mean your code works if you remove the AccuracyCalculator import?

[ZhuoxuanLiu]

Do you mean your code works if you remove the AccuracyCalculator import?

yes. In my env, pml can't work with umap.

[KevinMusgrave]

Wow, strange. I'll reopen the issue.

[domenicoMuscill0]

import numpy as np
from pytorch_metric_learning.utils.accuracy_calculator import AccuracyCalculator
import umap
    
if __name__ == '__main__':
    umapper = umap.UMAP()
    test_embeddings = np.random.normal(size=(100, 64))
    umap_embeddings = umapper.fit_transform(test_embeddings)

I have ran this code on ubuntu 22.04 LTS with umap-learn==0.5.3 and pytorch-metric-learning==2.2.0 and it works fine. Just some warnings for numba.jit decorator.