140 lines
4.3 KiB
Python
140 lines
4.3 KiB
Python
|
''' Define the sublayers in encoder/decoder layer
|
||
|
|
Derived From : https://github.com/jadore801120/attention-is-all-you-need-pytorch/blob/master/transformer/SubLayers.py
|
||
|
|
'''
|
||
|
|
|
||
|
|
import numpy as np
|
||
|
|
import torch
|
||
|
|
import torch.nn as nn
|
||
|
|
import torch.nn.functional as F
|
||
|
|
|
||
|
|
class ScaledDotProductAttention(nn.Module):
|
||
|
|
'''Scaled Dot-Product Attention
|
||
|
|
'''
|
||
|
|
|
||
|
|
def __init__(self, temperature):
|
||
|
|
'''
|
||
|
|
Initialize the model.
|
||
|
|
:param temperature: TODO ....
|
||
|
|
:param attn_dropout: Argument to dropout after softmax(QK)
|
||
|
|
'''
|
||
|
|
super().__init__()
|
||
|
|
self.temperature = temperature
|
||
|
|
|
||
|
|
def forward(self, q, k, v, mask=None):
|
||
|
|
'''
|
||
|
|
Callback Function:
|
||
|
|
:param q: The Query matrix.
|
||
|
|
:param k: The Key matrix.
|
||
|
|
:param v: The value matrix.
|
||
|
|
:param mask: The mask of the input.
|
||
|
|
:returns (output, attention): A tuple consisting of softmax(QK^T)V and softmax(QK^T)
|
||
|
|
'''
|
||
|
|
attn = torch.matmul(q / self.temperature, k.transpose(2, 3))
|
||
|
|
|
||
|
|
if mask is not None:
|
||
|
|
attn = attn.masked_fill(mask == 0, -1e9)
|
||
|
|
|
||
|
|
attn = F.softmax(attn, dim=-1)
|
||
|
|
output = torch.matmul(attn, v)
|
||
|
|
|
||
|
|
return output
|
||
|
|
|
||
|
|
|
||
|
|
class MultiHeadAttention(nn.Module):
|
||
|
|
''' Multi-Head Attention module '''
|
||
|
|
|
||
|
|
def __init__(self, n_head, d_model, d_k, d_v):
|
||
|
|
'''
|
||
|
|
Intialize the model.
|
||
|
|
:param n_head: Number of self-attention modules
|
||
|
|
:param d_model: Dimension of input/output of this layer
|
||
|
|
:param d_k: Dimension of each Key
|
||
|
|
:param d_v: Dimension of each Value
|
||
|
|
:param dropout:
|
||
|
|
'''
|
||
|
|
super().__init__()
|
||
|
|
|
||
|
|
self.n_head = n_head
|
||
|
|
self.d_k = d_k
|
||
|
|
self.d_v = d_v
|
||
|
|
|
||
|
|
self.w_qs = nn.Linear(d_model, n_head * d_k, bias=False)
|
||
|
|
self.w_ks = nn.Linear(d_model, n_head * d_k, bias=False)
|
||
|
|
self.w_vs = nn.Linear(d_model, n_head * d_v, bias=False)
|
||
|
|
self.fc = nn.Linear(n_head * d_v, d_model, bias=False)
|
||
|
|
|
||
|
|
self.attention = ScaledDotProductAttention(temperature=d_k ** 0.5)
|
||
|
|
|
||
|
|
self.layer_norm = nn.LayerNorm(d_model, eps=1e-6)
|
||
|
|
|
||
|
|
|
||
|
|
def forward(self, q, k, v, mask=None):
|
||
|
|
'''
|
||
|
|
Callback function.
|
||
|
|
:param q: The Query matrix.
|
||
|
|
:param k: The Key matrix.
|
||
|
|
:param v: The value matrix.
|
||
|
|
:param mask: The mask to use.
|
||
|
|
:returns (output, attention): A tuple consisting of network output and softmax(QK^T)
|
||
|
|
'''
|
||
|
|
d_k, d_v, n_head = self.d_k, self.d_v, self.n_head
|
||
|
|
sz_b_q = q.size(0)
|
||
|
|
sz_b, len_q, len_k, len_v = q.size(0), q.size(1), k.size(1), v.size(1)
|
||
|
|
|
||
|
|
residual = q
|
||
|
|
|
||
|
|
# Pass through the pre-attention projection: b x lq x (n*dv)
|
||
|
|
# Separate different heads: b x lq x n x dv
|
||
|
|
q = self.w_qs(q).view(sz_b_q, len_q, n_head, d_k)
|
||
|
|
k = self.w_ks(k).view(sz_b, len_k, n_head, d_k)
|
||
|
|
v = self.w_vs(v).view(sz_b, len_v, n_head, d_v)
|
||
|
|
|
||
|
|
# Transpose for attention dot product: b x n x lq x dv
|
||
|
|
q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
|
||
|
|
|
||
|
|
if mask is not None:
|
||
|
|
mask = mask.unsqueeze(1) # For head axis broadcasting.
|
||
|
|
|
||
|
|
q = self.attention(q, k, v, mask=mask)
|
||
|
|
|
||
|
|
# Transpose to move the head dimension back: b x lq x n x dv
|
||
|
|
# Combine the last two dimensions to concatenate all the heads together: b x lq x (n*dv)
|
||
|
|
q = q.transpose(1, 2).contiguous().view(sz_b, len_q, -1)
|
||
|
|
q = self.fc(q)
|
||
|
|
q += residual
|
||
|
|
|
||
|
|
q = self.layer_norm(q)
|
||
|
|
|
||
|
|
return q
|
||
|
|
|
||
|
|
|
||
|
|
class PositionwiseFeedForward(nn.Module):
|
||
|
|
''' A simple 2 layer with fully connected layer.
|
||
|
|
'''
|
||
|
|
|
||
|
|
def __init__(self, d_in, d_hid):
|
||
|
|
'''
|
||
|
|
Initialize the model.
|
||
|
|
:param d_in: Dimension of the input/output of the model.
|
||
|
|
:param d_hid: Dimension of the hidden layer.
|
||
|
|
:param dropout: Argument to the dropout layer.
|
||
|
|
'''
|
||
|
|
super().__init__()
|
||
|
|
self.w_1 = nn.Linear(d_in, d_hid) # position-wise
|
||
|
|
self.w_2 = nn.Linear(d_hid, d_in) # position-wise
|
||
|
|
self.layer_norm = nn.LayerNorm(d_in, eps=1e-6)
|
||
|
|
|
||
|
|
def forward(self, x):
|
||
|
|
'''
|
||
|
|
Callback function.
|
||
|
|
:param x: The input to the function.
|
||
|
|
:returns torch.array: An output of the same dimension as the input.
|
||
|
|
'''
|
||
|
|
residual = x
|
||
|
|
|
||
|
|
x = self.w_2(F.relu(self.w_1(x)))
|
||
|
|
x += residual
|
||
|
|
|
||
|
|
x = self.layer_norm(x)
|
||
|
|
|
||
|
|
return x
|