File size: 7,809 Bytes
659e74f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
import math
import torch 
import torch.nn.functional as F 
from torch import nn
from torch.nn.modules.normalization import LayerNorm


class DilatedMultiheadSelfAttentionWithRelativePositionalEmbedding(nn.Module):
    def __init__(self, dmodel, num_heads, dropout=0, Er_provided=False, attn_len=5):
        super(DilatedMultiheadSelfAttentionWithRelativePositionalEmbedding, self).__init__()
        self.attn_len = attn_len
        self.dmodel = dmodel
        self.num_heads = num_heads
        self.head_dim = dmodel // num_heads
        assert self.head_dim * num_heads == dmodel, "embed_dim must be divisible by num_heads"

        self.key = nn.Linear(dmodel, dmodel)
        self.value = nn.Linear(dmodel, dmodel)
        self.query = nn.Linear(dmodel, dmodel)
        self.dropout = nn.Dropout(dropout)
        self.Er_provided = Er_provided
        
        if not Er_provided:
            self.Er = nn.Parameter(torch.randn(num_heads, self.head_dim, attn_len))


    def forward(self, query, key, value, layer=0):
        #query, key, and value: (batch, time, dmodel), float tensor

        batch, time, d_model = query.shape

        q = self.query(query).reshape(batch, time, self.num_heads, 1, self.head_dim).transpose(1, 2)  #(batch, num_head, time, 1, head_dim)
        k = self.key(key).reshape(batch, time, self.num_heads, 1, self.head_dim).transpose(1, 2)  #(batch, num_head, time, 1, head_dim)
        v = self.value(value).reshape(batch, time, self.num_heads, 1, self.head_dim).transpose(1, 2)  #(batch, num_head, time, 1, head_dim)

        k = torch.cat(
                        (
                        self.kv_roll(k[:, 0: 4], layer, padding_value=0, shift=0),
                        self.kv_roll(k[:, 4: 5], layer, padding_value=0, shift=-2),
                        self.kv_roll(k[:, 5: 6], layer, padding_value=0, shift=-1),
                        self.kv_roll(k[:, 6: 7], layer, padding_value=0, shift=1),
                        self.kv_roll(k[:, 6: 7], layer, padding_value=0, shift=2)   #This should be k[:, 7: 8]. We had this bug during training so we keep it to fit the checkpoints.
                        ),
                    dim=1
                    )   #we define 4 symmetrical heads and 4 skewed heads

        v = torch.cat(
                        (
                        self.kv_roll(v[:, 0: 4], layer, padding_value=0, shift=0),
                        self.kv_roll(v[:, 4: 5], layer, padding_value=0, shift=-2),
                        self.kv_roll(v[:, 5: 6], layer, padding_value=0, shift=-1),
                        self.kv_roll(v[:, 6: 7], layer, padding_value=0, shift=1),
                        self.kv_roll(v[:, 7: 8], layer, padding_value=0, shift=2)
                        ),
                    dim=1
                    )   #we define 4 symmetrical heads and 4 skewed heads
        
        Er_t = self.Er.unsqueeze(1).unsqueeze(0)  #(1, num_head, 1, head_dim, attn_len)

        qk = torch.matmul(q, k.transpose(-2, -1))
        attn_mask = torch.zeros_like(qk).masked_fill_((qk==0), float('-inf'))
        attn = (qk + torch.matmul(q, Er_t)) / math.sqrt(self.head_dim)
        attn = F.softmax(attn + attn_mask, dim=-1)

        out = torch.matmul(attn, v) #(batch, num_head, time, 1, head_dim)
        out = out.squeeze(-2).transpose(1, 2).reshape(batch, time, d_model)

        return self.dropout(out), attn
        


    def kv_roll(self, tensor, layer, padding_value=0, shift=1):
        #tensor: (batch, num_head, time, 1, head_dim)
        batch, num_head, time, _, head_dim = tensor.shape

        tensor = F.pad(tensor, (0, 0, 0, 0, (2**layer)*(self.attn_len//2), (2**layer)*(self.attn_len//2)), mode='constant', value=padding_value) 
        #(batch, num_head, time+(2**layer)*(self.attn_len//2), 1, head_dim)

        tensor = torch.cat([torch.roll(tensor, shifts=-i*(2**layer), dims=2) for i in range(shift, self.attn_len+shift)], dim=-2) 
        #(batch, num_head, time+(2**layer)*(self.attn_len//2), attn_len, head_dim)

        return tensor[:, :, :time, :, :]    #(batch, num_head, time, attn_len, head_dim)




class DilatedTransformerLayer(nn.Module):
    def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1, Er_provided=False, attn_len=5, norm_first=False, layer_norm_eps=1e-5):
        super(DilatedTransformerLayer, self).__init__()
        self.self_attn = DilatedMultiheadSelfAttentionWithRelativePositionalEmbedding(d_model, nhead, dropout, Er_provided, attn_len)
        # Implementation of Feedforward model
        self.linear1 = nn.Linear(d_model, dim_feedforward)
        self.dropout = nn.Dropout(dropout)
        self.linear2 = nn.Linear(dim_feedforward, d_model)

        self.norm_first = norm_first
        self.norm1 = LayerNorm(d_model, eps=layer_norm_eps)
        self.norm2 = LayerNorm(d_model, eps=layer_norm_eps)
        self.dropout1 = nn.Dropout(dropout)
        self.dropout2 = nn.Dropout(dropout)

        self.activation = F.gelu


    def forward(self, x, layer=0):
        #x: (batch, time, dmodel)
        if self.norm_first:
            x_ = self._sa_block(self.norm1(x), layer)[0]
            x = x + x_
            x = x + self._ff_block(self.norm2(x))
        else:
            x_ = self._sa_block(x, layer)[0]
            x = self.norm1(x + x_)
            x = self.norm2(x + self._ff_block(x))
        return x, x_


    def inference(self, x, layer=0):
        #x: (batch, time, dmodel)
        if self.norm_first:
            x_, attn = self._sa_block(self.norm1(x), layer)
            x = x + x_
            x = x + self._ff_block(self.norm2(x))
        else:
            x_, attn = self._sa_block(x, layer)
            x = self.norm1(x + x_)
            x = self.norm2(x + self._ff_block(x))


        attn = attn.squeeze(-2) #batch, num_head, time, attn_len
        batch, num_head, time, attn_len = attn.shape
        padded_attn_len = (attn_len-1) * (2**layer) + 1
        tmp_output = torch.zeros(batch, num_head, time, padded_attn_len, device=x.device)
        for i, j in enumerate(range(0, padded_attn_len, 2**layer)):
            tmp_output[:, :, :, j] = attn[:, :, :, i]

        attn = torch.zeros(batch, num_head, time, time+(padded_attn_len-1)*2, device=x.device)
        for i in range(time):
            attn[:, :, i, i: i+padded_attn_len] = tmp_output[:, :, i]

        center = (padded_attn_len-1)
        attn = torch.cat(
                            [
                                attn[:, 0: 4, :,  center - (2**layer) * 2:  center - (2**layer) * 2 + time],
                                attn[:, 4: 5, :,  center - (2**layer) * 1:  center - (2**layer) * 1 + time],
                                attn[:, 5: 6, :,  center - (2**layer) * 0:  center - (2**layer) * 0 + time],
                                attn[:, 6: 7, :,  center - (2**layer) * 3:  center - (2**layer) * 3 + time],
                                attn[:, 7: 8, :,  center - (2**layer) * 4:  center - (2**layer) * 4 + time]
                            ],
                            dim=1
                        )   #restore the square attention matrix from dilated self-attention

        return x, x_, attn


    # self-attention block
    def _sa_block(self, x, layer=0):
        x, attn = self.self_attn(x, x, x, layer)
        return self.dropout1(x), attn


    # feed forward block
    def _ff_block(self, x):
        x = self.linear2(self.dropout(self.activation(self.linear1(x))))
        return self.dropout2(x)




if __name__ == '__main__':
    BATCH=1
    TIME=9
    DMODEL=8
    N_HEAD=4
    ATTN_LEN=5
    LAYER=1

    x = torch.ones(BATCH, TIME, DMODEL)

    model = DilatedMultiheadSelfAttentionWithRelativePositionalEmbedding(dmodel=DMODEL, num_heads=N_HEAD, attn_len=ATTN_LEN)

    output, attn = model(x, x, x, layer=LAYER)
    print(attn[0, 0, :, :, :])