File size: 7,956 Bytes
4bf9661
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
import torch
from torch import einsum, nn
import torch.nn.functional as F
from einops import rearrange, repeat

# helper functions

def exists(val):
    return val is not None

def default(val, d):
    return val if exists(val) else d

# normalization
# they use layernorm without bias, something that pytorch does not offer


class LayerNorm(nn.Module):
    def __init__(self, dim):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(dim))
        self.register_buffer("bias", torch.zeros(dim))

    def forward(self, x):
        return F.layer_norm(x, x.shape[-1:], self.weight, self.bias)

# residual


class Residual(nn.Module):
    def __init__(self, fn):
        super().__init__()
        self.fn = fn

    def forward(self, x, *args, **kwargs):
        return self.fn(x, *args, **kwargs) + x


# rotary positional embedding
# https://arxiv.org/abs/2104.09864


class RotaryEmbedding(nn.Module):
    def __init__(self, dim):
        super().__init__()
        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq)

    def forward(self, max_seq_len, *, device):
        seq = torch.arange(max_seq_len, device=device, dtype=self.inv_freq.dtype)
        freqs = einsum("i , j -> i j", seq, self.inv_freq)
        return torch.cat((freqs, freqs), dim=-1)


def rotate_half(x):
    x = rearrange(x, "... (j d) -> ... j d", j=2)
    x1, x2 = x.unbind(dim=-2)
    return torch.cat((-x2, x1), dim=-1)


def apply_rotary_pos_emb(pos, t):
    return (t * pos.cos()) + (rotate_half(t) * pos.sin())


# classic Noam Shazeer paper, except here they use SwiGLU instead of the more popular GEGLU for gating the feedforward
# https://arxiv.org/abs/2002.05202


class SwiGLU(nn.Module):
    def forward(self, x):
        x, gate = x.chunk(2, dim=-1)
        return F.silu(gate) * x


# parallel attention and feedforward with residual
# discovered by Wang et al + EleutherAI from GPT-J fame

class ParallelTransformerBlock(nn.Module):
    def __init__(self, dim, dim_head=64, heads=8, ff_mult=4):
        super().__init__()
        self.norm = LayerNorm(dim)

        attn_inner_dim = dim_head * heads
        ff_inner_dim = dim * ff_mult
        self.fused_dims = (attn_inner_dim, dim_head, dim_head, (ff_inner_dim * 2))

        self.heads = heads
        self.scale = dim_head**-0.5
        self.rotary_emb = RotaryEmbedding(dim_head)

        self.fused_attn_ff_proj = nn.Linear(dim, sum(self.fused_dims), bias=False)
        self.attn_out = nn.Linear(attn_inner_dim, dim, bias=False)

        self.ff_out = nn.Sequential(
            SwiGLU(),
            nn.Linear(ff_inner_dim, dim, bias=False)
        )

        self.register_buffer("pos_emb", None, persistent=False)


    def get_rotary_embedding(self, n, device):
        if self.pos_emb is not None and self.pos_emb.shape[-2] >= n:
            return self.pos_emb[:n]

        pos_emb = self.rotary_emb(n, device=device)
        self.register_buffer("pos_emb", pos_emb, persistent=False)
        return pos_emb

    def forward(self, x, attn_mask=None):
        """
        einstein notation
        b - batch
        h - heads
        n, i, j - sequence length (base sequence length, source, target)
        d - feature dimension
        """

        n, device, h = x.shape[1], x.device, self.heads

        # pre layernorm

        x = self.norm(x)

        # attention queries, keys, values, and feedforward inner

        q, k, v, ff = self.fused_attn_ff_proj(x).split(self.fused_dims, dim=-1)

        # split heads
        # they use multi-query single-key-value attention, yet another Noam Shazeer paper
        # they found no performance loss past a certain scale, and more efficient decoding obviously
        # https://arxiv.org/abs/1911.02150

        q = rearrange(q, "b n (h d) -> b h n d", h=h)

        # rotary embeddings

        positions = self.get_rotary_embedding(n, device)
        q, k = map(lambda t: apply_rotary_pos_emb(positions, t), (q, k))

        # scale

        q = q * self.scale

        # similarity

        sim = einsum("b h i d, b j d -> b h i j", q, k)


        # extra attention mask - for masking out attention from text CLS token to padding

        if exists(attn_mask):
            attn_mask = rearrange(attn_mask, 'b i j -> b 1 i j')
            sim = sim.masked_fill(~attn_mask, -torch.finfo(sim.dtype).max)

        # attention

        sim = sim - sim.amax(dim=-1, keepdim=True).detach()
        attn = sim.softmax(dim=-1)

        # aggregate values

        out = einsum("b h i j, b j d -> b h i d", attn, v)

        # merge heads

        out = rearrange(out, "b h n d -> b n (h d)")
        return self.attn_out(out) + self.ff_out(ff)

# cross attention - using multi-query + one-headed key / values as in PaLM w/ optional parallel feedforward

class CrossAttention(nn.Module):
    def __init__(
        self,
        dim,
        *,
        context_dim=None,
        dim_head=64,
        heads=12,
        parallel_ff=False,
        ff_mult=4,
        norm_context=False
    ):
        super().__init__()
        self.heads = heads
        self.scale = dim_head ** -0.5
        inner_dim = heads * dim_head
        context_dim = default(context_dim, dim)

        self.norm = LayerNorm(dim)
        self.context_norm = LayerNorm(context_dim) if norm_context else nn.Identity()

        self.to_q = nn.Linear(dim, inner_dim, bias=False)
        self.to_kv = nn.Linear(context_dim, dim_head * 2, bias=False)
        self.to_out = nn.Linear(inner_dim, dim, bias=False)

        # whether to have parallel feedforward

        ff_inner_dim = ff_mult * dim

        self.ff = nn.Sequential(
            nn.Linear(dim, ff_inner_dim * 2, bias=False),
            SwiGLU(),
            nn.Linear(ff_inner_dim, dim, bias=False)
        ) if parallel_ff else None

    def forward(self, x, context, mask):
        """
        einstein notation
        b - batch
        h - heads
        n, i, j - sequence length (base sequence length, source, target)
        d - feature dimension
        """

        # pre-layernorm, for queries and context

        x = self.norm(x)
        context = self.context_norm(context)

        # get queries

        q = self.to_q(x)
        q = rearrange(q, 'b n (h d) -> b h n d', h = self.heads)

        # scale

        q = q * self.scale

        # get key / values

        k, v = self.to_kv(context).chunk(2, dim=-1)

        # query / key similarity

        sim = einsum('b h i d, b j d -> b h i j', q, k)

        # attention
        mask = mask.unsqueeze(1).repeat(1,self.heads,1,1)
        sim = sim + mask  # context mask
        sim = sim - sim.amax(dim=-1, keepdim=True)
        attn = sim.softmax(dim=-1)

        # aggregate

        out = einsum('b h i j, b j d -> b h i d', attn, v)

        # merge and combine heads

        out = rearrange(out, 'b h n d -> b n (h d)')
        out = self.to_out(out)

        # add parallel feedforward (for multimodal layers)

        if exists(self.ff):
            out = out + self.ff(x)

        return out


class Cross_model(nn.Module):
    def __init__(
        self,
        dim=512,
        layer_num=4,
        dim_head=64,
        heads=8,
        ff_mult=4
    ):
        super().__init__()

        self.layers = nn.ModuleList([])


        for ind in range(layer_num):
            self.layers.append(nn.ModuleList([
                Residual(CrossAttention(dim=dim, dim_head=dim_head, heads=heads, parallel_ff=True, ff_mult=ff_mult)),
                Residual(ParallelTransformerBlock(dim=dim, dim_head=dim_head, heads=heads, ff_mult=ff_mult))
            ]))

    def forward(
        self,
        query_tokens,
        context_tokens,
        mask
    ):

        for cross_attn, self_attn_ff in self.layers:
            query_tokens = cross_attn(query_tokens, context_tokens,mask)
            query_tokens = self_attn_ff(query_tokens)

        return query_tokens