Learned queries attend to the visual patches through
morie_alammar_sdp_attention, with an optional output projection.
Examples
V <- c(1, 2, 3, 4, 5, 6, 7, 8)
morie_kamath_q_former(V, V)
#> $Z
#> [,1]
#> [1,] 7.420708
#> [2,] 7.843483
#> [3,] 7.947604
#> [4,] 7.981343
#> [5,] 7.993216
#> [6,] 7.997515
#> [7,] 7.999087
#> [8,] 7.999664
#>
#> $attention
#> [,1] [,2] [,3] [,4] [,5]
#> [1,] 5.766128e-04 1.567396e-03 4.260624e-03 1.158158e-02 3.148199e-02
#> [2,] 7.189936e-07 5.312684e-06 3.925572e-05 2.900627e-04 2.143290e-03
#> [3,] 7.205047e-10 1.447172e-08 2.906723e-07 5.838310e-06 1.172656e-04
#> [4,] 6.787758e-13 3.705991e-11 2.023402e-09 1.104740e-07 6.031677e-06
#> [5,] 6.262633e-16 9.294572e-14 1.379437e-11 2.047266e-09 3.038412e-07
#> [6,] 5.735271e-19 2.313773e-16 9.334428e-14 3.765777e-11 1.519223e-08
#> [7,] 5.238105e-22 5.744279e-19 6.299367e-16 6.908095e-13 7.575646e-10
#> [8,] 4.779289e-25 1.424686e-21 4.246929e-18 1.265992e-14 3.773868e-11
#> [,6] [,7] [,8]
#> [1,] 8.557692e-02 0.2326221940 0.6323327
#> [2,] 1.583689e-02 0.1170196575 0.8646648
#> [3,] 2.355342e-03 0.0473083162 0.9502129
#> [4,] 3.293184e-04 0.0179801763 0.9816844
#> [5,] 4.509403e-05 0.0066925471 0.9932621
#> [6,] 6.128982e-06 0.0024726080 0.9975212
#> [7,] 8.307705e-07 0.0009110504 0.9990881
#> [8,] 1.124974e-07 0.0003353501 0.9996645
#>
#> $n_queries
#> [1] 8
#>
#> $n_patches
#> [1] 8
#>
#> $compression
#> [1] 1
#>
#> $estimate
#> [1] 7.420708
#>
#> $n
#> [1] 8
#>
#> $method
#> [1] "Q-Former cross-attention (reused attention core)"
#>