File size: 7,997 Bytes
b0b3b00
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Based on: https://github.com/lucidrains/flamingo-pytorch
"""

import torch
from einops import rearrange, repeat
from torch import einsum, nn
from einops_exts import rearrange_many

def exists(val):
    return val is not None


def FeedForward(
    dim,
    mult=4,
    use_ft_layernorm=False,
    enable_init_network_params=False,
    initializer_range=0.02,
):
    inner_dim = int(dim * mult)
    net = nn.Sequential(
        nn.LayerNorm(dim),
        nn.Linear(dim, inner_dim, bias=False),
        nn.GELU(),
        nn.Linear(inner_dim, dim, bias=False),
    )

    if use_ft_layernorm and enable_init_network_params:
        # only use_ft_layernorm is on and enalbe_init_network_params
        # then start the initialization
        net[0].weight.data.normal_(mean=0.0, std=initializer_range)
        net[0].bias.data.zero_()
        net[1].weight.data.normal_(mean=0.0, std=initializer_range)
        net[3].weight.data.normal_(mean=0.0, std=initializer_range)
    return net


# gated cross attention
class MaskedCrossAttention(nn.Module):
    def __init__(
        self,
        *,
        dim,
        dim_visual,
        dim_head=64,
        heads=8,
        only_attend_immediate_media=True,
        use_ft_layernorm=False,
        use_ft_flash_attention=False,
        enable_init_network_params=False,
        initializer_range=0.02,
    ):
        super().__init__()
        self.scale = dim_head**-0.5
        self.heads = heads
        self.use_ft_flash_attention = False
        self.initializer_range = initializer_range
        inner_dim = dim_head * heads

        self.norm = nn.LayerNorm(dim)

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

        # whether for text to only attend to immediate preceding image, or all previous images
        self.only_attend_immediate_media = only_attend_immediate_media

        if enable_init_network_params:
            self.apply(self._init_weights)

    def _init_weights(self, module):
        if isinstance(module, nn.Linear):
            # Slightly different from the TF version which uses truncated_normal for initialization
            # cf https://github.com/pytorch/pytorch/pull/5617
            module.weight.data.normal_(mean=0.0, std=self.initializer_range)
            if module.bias is not None:
                module.bias.data.zero_()

        elif isinstance(module, nn.LayerNorm):
            module.bias.data.zero_()
            module.weight.data.fill_(1.0)

    def forward(self, x, media, media_locations=None, use_cached_media=False, image_mask=None):
        """
        Args:
            x (torch.Tensor): text features
                shape (B, T_txt, D_txt)
            media (torch.Tensor): image features
                shape (B, T_img, n, D_img) where n is the dim of the latents
            media_locations: boolean mask identifying the media tokens in x
                shape (B, T_txt)
            use_cached_media: bool
                If true, treat all of x as if they occur after the last media
                registered in media_locations. T_txt does not need to exactly
                equal media_locations.shape[1] in this case
        """

        if not use_cached_media:
            assert media_locations.shape[1] == x.shape[1], (
                f"media_location.shape is {media_locations.shape} but x.shape is"
                f" {x.shape}"
            )

        T_txt = x.shape[1]
        _, T_img, n = media.shape[:3]
        h = self.heads

        x = self.norm(x.contiguous())
        q = self.to_q(x)
        media = rearrange(media, "b t n d -> b (t n) d")

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

        if exists(media_locations):
            media_time = torch.arange(T_img, device=x.device) + 1

            if use_cached_media:
                # text time is set to the last cached media location
                text_time = repeat(
                    torch.count_nonzero(media_locations, dim=1),
                    "b -> b i",
                    i=T_txt,
                )
            else:
                # at each boolean of True, increment the time counter (relative to media time)
                text_time = media_locations.cumsum(dim=-1)

            # text time must equal media time if only attending to most immediate image
            # otherwise, as long as text time is greater than media time (if attending to all previous images / media)
            mask_op = torch.eq if self.only_attend_immediate_media else torch.ge
            text_to_media_mask = mask_op(
                rearrange(text_time, "b i -> b 1 i 1"),
                repeat(media_time, "j -> 1 1 1 (j n)", n=n),
            )

            if self.only_attend_immediate_media:
                # any text without a preceding media needs to have attention zeroed out
                text_without_media_mask = text_time == 0
                text_without_media_mask = rearrange(
                    text_without_media_mask, "b i -> b 1 i 1"
                )

        
        q, k, v = rearrange_many((q, k, v), "b n (h d) -> b h n d", h=h)
        q = q * self.scale
        sim = einsum("... i d, ... j d -> ... i j", q, k)

        if exists(image_mask):
            image_mask = image_mask.unsqueeze(1).unsqueeze(1).bool()
            image_mask = image_mask.repeat_interleave(int(sim.shape[3] / image_mask.shape[3]), dim=-1)
            sim = sim.masked_fill(~image_mask, -torch.finfo(sim.dtype).max)
        # if exists(media_locations):
        #     sim = sim.masked_fill(~text_to_media_mask, -torch.finfo(sim.dtype).max)

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

        if exists(media_locations) and self.only_attend_immediate_media:
            # any text without a preceding media needs to have attention zeroed out
            attn = attn.masked_fill(text_without_media_mask, 0.0)

        out = einsum("... i j, ... j d -> ... i d", attn, v)
        out = rearrange(out, "b h n d -> b n (h d)")
        
        return self.to_out(out)



class GatedCrossAttentionBlock(nn.Module):
    def __init__(
        self,
        *,
        dim,
        dim_visual,
        dim_head=64,
        heads=12,
        ff_mult=1,
        only_attend_immediate_media=True,
        use_ft_layernorm=False,
        use_ft_flash_attention=False,
        enable_init_network_params=False,
        initializer_range=0.02,
        gradient_checkpointing=False,
    ):
        super().__init__()
        self.attn = MaskedCrossAttention(
            dim=dim,
            dim_visual=dim_visual,
            dim_head=dim_head,
            heads=heads,
            only_attend_immediate_media=only_attend_immediate_media,
            use_ft_flash_attention=use_ft_flash_attention,
            use_ft_layernorm=use_ft_layernorm,
            enable_init_network_params=enable_init_network_params,
            initializer_range=initializer_range,
        )
        self.attn_gate = nn.Parameter(torch.zeros(dim))

        self.ff = FeedForward(dim, mult=ff_mult)
        self.ff_gate = nn.Parameter(torch.zeros(dim))

        self.gradient_checkpointing = gradient_checkpointing

    def forward(
        self,
        x,
        media,
        media_locations=None,
        use_cached_media=False,
        image_mask=None,
    ):
        
        flag = torch.sum(media_locations, dim=-1)
        flag = torch.where(flag > 0.0, 1.0, 0.0)
        flag = flag.unsqueeze(1).unsqueeze(1).to(torch.bfloat16)
        x = (
            flag
            * self.attn(
                x,
                media,
                media_locations=media_locations,
                use_cached_media=use_cached_media,
                image_mask=image_mask,
            )
            * self.attn_gate.tanh()
            + x
        )

        x = flag * self.ff(x) * self.ff_gate.tanh() + x

        return x