I have a few lines of code where I have a torch tensor (200.000 x 512) and I want to calculate the cosine distance between each of those embeddings.
Right now, I'm doing:
similarity = (image_features.cpu() @ image_features.cpu().T).squeeze()
But it takes too long. Maybe matlab is much faster. Does anyone can tell me how to do it there? (Convert my tensor to .mat file + code)
You can convert your torch tensor into a numpy array,
then save that to a .mat with scipy's savemat().
from scipy.io import savemat
# to numpy array
arr = tensor.numpy()
# to .mat, needs second input to be a dict type
savemat("filename.mat", {'a': arr})
However, MATLAB isn't always faster than Python (Numpy); especially when doing matrix multiplications, since Numpy itself is also highly optimized. So I believe that trying the same matrix multiplication on MATLAB won't speed things up. In fact, it might be faster to do the calculation in pytorch if you have a GPU available.
Perhaps a more relevant question here is how to calculate the cosine similarity of large datasets?
