File size: 2,732 Bytes
64e1ee8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
import torch
from torch import nn
from torch.nn import functional as F
import math

class SelfAttention(nn.Module):
    def __init__(self, n_heads, d_embed, in_proj_bias=True, out_proj_bias=True):
        super().__init__()
        self.in_proj = nn.Linear(d_embed, 3 * d_embed, bias=in_proj_bias)
        self.out_proj = nn.Linear(d_embed, d_embed, bias=out_proj_bias)
        self.n_heads = n_heads
        self.d_head = d_embed // n_heads

    def forward(self, x, causal_mask=False):
        input_shape = x.shape 
        batch_size, sequence_length, d_embed = input_shape 
        interim_shape = (batch_size, sequence_length, self.n_heads, self.d_head) 

        q, k, v = self.in_proj(x).chunk(3, dim=-1)        
        q = q.view(interim_shape).transpose(1, 2)
        k = k.view(interim_shape).transpose(1, 2)
        v = v.view(interim_shape).transpose(1, 2)

        weight = q @ k.transpose(-1, -2)
        
        if causal_mask:
            mask = torch.ones_like(weight, dtype=torch.bool).triu(1) 
            weight.masked_fill_(mask, -torch.inf) 
        
        weight /= math.sqrt(self.d_head) 

        weight = F.softmax(weight, dim=-1) 
        output = weight @ v
        output = output.transpose(1, 2) 

        output = output.reshape(input_shape) 
        output = self.out_proj(output) 
        return output

class CrossAttention(nn.Module):
    def __init__(self, n_heads, d_embed, d_cross, in_proj_bias=True, out_proj_bias=True):
        super().__init__()
        self.q_proj   = nn.Linear(d_embed, d_embed, bias=in_proj_bias)
        self.k_proj   = nn.Linear(d_cross, d_embed, bias=in_proj_bias)
        self.v_proj   = nn.Linear(d_cross, d_embed, bias=in_proj_bias)
        self.out_proj = nn.Linear(d_embed, d_embed, bias=out_proj_bias)
        self.n_heads = n_heads
        self.d_head = d_embed // n_heads
    
    def forward(self, x, y):
        input_shape = x.shape
        batch_size, sequence_length, d_embed = input_shape
        interim_shape = (batch_size, -1, self.n_heads, self.d_head)
        
        q = self.q_proj(x)
        k = self.k_proj(y)
        v = self.v_proj(y)

        q = q.view(interim_shape).transpose(1, 2) 
        k = k.view(interim_shape).transpose(1, 2) 
        v = v.view(interim_shape).transpose(1, 2) 
        
        weight = q @ k.transpose(-1, -2)
        
        weight /= math.sqrt(self.d_head)
        
        weight = F.softmax(weight, dim=-1)
        
        output = weight @ v
        
        output = output.transpose(1, 2).contiguous()
        
        output = output.view(input_shape)
        
        output = self.out_proj(output)

        return output