In a variety of different fingerprint kernels, x1 and x2 are compared with something like
x1[-1] <something> x2[-1]
which only compares the last element in the kernel matrix, and does not do a pairwise comparison. This causes many kernels to not implement the function claimed in their docstring. Specifically, I think the following kernels are wrong:
- Braun-Blanquet: the max(x1, x2) in the denominator is not computed correctly here:
|
denom = torch.max(x1_norm[-1], x2_norm[-1]) |
- Faith: number of matching 0s (
d) is only computed with the last element:
|
d = torch.sum((x1[-1] == 0) & (x2[-1] == 0), dim=-1, keepdims=True) |
- Intersection kernel: same problem as Faith
- Rand: same problem as Faith
- Rogers-Tanimoto: same problem as Faith
Here is a minimal example showing the problem:
import torch
from gauche.kernels.fingerprint_kernels.rand_kernel import batch_rand_sim
t = torch.as_tensor([[1, 1, 0., 0.], [1., 0., 1., 0.]])
print(batch_rand_sim(t, t)[0, 1])
print(batch_rand_sim(t[:1], t[1:]))
Output:
tensor(0.7500, dtype=torch.float64)
tensor([[0.5000]], dtype=torch.float64)
What this shows: if I compute a whole kernel matrix and take element [0, -1], it gives a different result than directly computing element [0, -1]. This is because d is computed just using the last element in each batch, rather than computed separately for each pair.
I think this was not caught in the tests because they just tested 1xD matrices. I suspect that a lot of the "batch" similarities also don't work for batches of matrices.
If you can't fix this promptly, I suggest taking these methods out of Gauche and re-releasing. You can add corrected methods back in later.
In a variety of different fingerprint kernels,
x1andx2are compared with something likewhich only compares the last element in the kernel matrix, and does not do a pairwise comparison. This causes many kernels to not implement the function claimed in their docstring. Specifically, I think the following kernels are wrong:
gauche/gauche/kernels/fingerprint_kernels/braun_blanquet_kernel.py
Line 37 in 3380401
d) is only computed with the last element:gauche/gauche/kernels/fingerprint_kernels/faith_kernel.py
Line 35 in 3380401
Here is a minimal example showing the problem:
Output:
What this shows: if I compute a whole kernel matrix and take element [0, -1], it gives a different result than directly computing element [0, -1]. This is because
dis computed just using the last element in each batch, rather than computed separately for each pair.I think this was not caught in the tests because they just tested
1xDmatrices. I suspect that a lot of the "batch" similarities also don't work for batches of matrices.If you can't fix this promptly, I suggest taking these methods out of Gauche and re-releasing. You can add corrected methods back in later.