pytorch-grad-cam
pytorch-grad-cam copied to clipboard
eigen cam generate wrong results
The Eigen cam used the SVD from NumPy. Depending on the NumPy optimizer, the sign of eigenvectors could be negative. It will lead the algorithm to output the non-attention region instead of the attention region. I found this issue in MKL.
Did you fix it yourself?
SOLVED: I just use SciPy SVD to fix it
@jacobgil Could you update get_2d_projection function? NumPy SVD to SciPy SVD?
Did you fix it yourself?
SOLVED: I just use SciPy SVD to fix it
@jacobgil Could you update get_2d_projection function? NumPy SVD to SciPy SVD?
Hello, I use scipy, but the result seems same to numpy, which part of the code has been changed?