103 lines
2.7 KiB
Julia
103 lines
2.7 KiB
Julia
# using LuxCore
|
||
using Random: AbstractRNG
|
||
using Lux
|
||
|
||
"""
|
||
# Fields
|
||
- `embed_dim`: Queue vector length. `q_len`
|
||
- `kdim::Int`: Key vector length.
|
||
- `vdim::Int`: Value vector length
|
||
- `num_heads`: Number of heads.
|
||
"""
|
||
struct MultiheadAttention{F} <: LuxCore.AbstractLuxLayer
|
||
embed_dim::Int
|
||
kdim::Int
|
||
vdim::Int
|
||
num_heads::Int
|
||
init_weight::F
|
||
end
|
||
|
||
"""
|
||
MultiheadAttention(embed_dim::Int, num_heads::Int; init_weight=glorot_uniform, kw...)
|
||
|
||
Constructor.
|
||
|
||
# Arguments
|
||
- `embed_dim::Int`
|
||
- `num_heads::Int`
|
||
|
||
## Keyword Arguments
|
||
- `init_weight`: weight initialzer (rng generator)
|
||
- `kdim::Int`: Default: `embed_dim`
|
||
- `vdim::Int`: Default: `embed_dim`
|
||
|
||
# Parameters and states
|
||
|
||
## Parameters
|
||
|
||
"""
|
||
function MultiheadAttention(
|
||
embed_dim::Int,
|
||
num_heads::Int;
|
||
init_weight = glorot_uniform,
|
||
kw...,
|
||
)
|
||
MultiheadAttention{typeof(init_weight)}(
|
||
embed_dim,
|
||
haskey(kw, :kdim) ? kw[:kdim] : embed_dim,
|
||
haskey(kw, :vdim) ? kw[:vdim] : embed_dim,
|
||
num_heads,
|
||
init_weight,
|
||
)
|
||
end
|
||
|
||
function LuxCore.initialparameters(rng::AbstractRNG, l::MultiheadAttention)
|
||
# see the original paper for weight dimensions (note that q,k,v weights have `num_heads` of matrices)
|
||
(
|
||
weight_q = l.init_weight(rng, l.embed_dim * l.num_heads, l.embed_dim),
|
||
weight_k = l.init_weight(rng, l.embed_dim * l.num_heads, l.kdim),
|
||
weight_v = l.init_weight(rng, l.embed_dim * l.num_heads, l.vdim),
|
||
weight_o = l.init_weight(rng, 10), # TODO
|
||
)
|
||
end
|
||
|
||
function LuxCore.initialstates(::AbstractRNG, ::MultiheadAttention)
|
||
NamedTuple()
|
||
end
|
||
|
||
function (l::MultiheadAttention)(
|
||
x::NamedTuple,
|
||
ps,
|
||
st::NamedTuple,
|
||
)
|
||
if size(x.q, 1) != l.embed_dim
|
||
ArgumentError(
|
||
"Length of queue must match the layer's embed_dim: size(q)[1] = $(size(x.q, 1)), embed_dim = $(l.embed_dim)",
|
||
) |> throw
|
||
end
|
||
if size(x.k, 1) != l.kdim
|
||
ArgumentError(
|
||
"Length of key must match the layer's kdim: size(k)[1] = $(size(x.k, 1)), kdim = $(l.kdim)",
|
||
) |> throw
|
||
end
|
||
if size(x.v, 1) != l.vdim
|
||
ArgumentError(
|
||
"Length of value must match the layer's vdim: size(v)[1] = $(size(x.v, 1)), vdim = $(l.vdim)",
|
||
) |> throw
|
||
end
|
||
# TODO
|
||
|
||
# qk_dim, v_dim is divisible by num_heads. qk_dim = embed_dim * num_heads
|
||
# [q] = (qk_dim, q_len, batch_size...)
|
||
q = ps.weight_q * x.q
|
||
# [k] = (qk_dim, kv_len, batch_size...)
|
||
k = ps.weight_k * x.k
|
||
# [v] = (v_dim, kv_len, batch_size...)
|
||
v = ps.weight_v * x.v # TODO: dimension?? 2025-01-28T21:59:56+09:00
|
||
# [y] = (v_dim, q_len, batch_size...)
|
||
y, α = dot_product_attention(q, k, v; nheads = l.num_heads)
|
||
end
|
||
|
||
# struct TransformerEncoder <: LuxCore.AbstractLuxContainerLayer{}
|
||
# end
|
||
#
|