File size: 7,734 Bytes
2420b7f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from torch import nn
import torch.nn.functional as nnf
from transformers import GPT2Tokenizer, GPT2LMHeadModel
import torch
from typing import Tuple, List, Union, Optional
import numpy as np


N = type(None)
V = np.array
ARRAY = np.ndarray
ARRAYS = Union[Tuple[ARRAY, ...], List[ARRAY]]
VS = Union[Tuple[V, ...], List[V]]
VN = Union[V, N]
VNS = Union[VS, N]
T = torch.Tensor
TS = Union[Tuple[T, ...], List[T]]
TN = Optional[T]
TNS = Union[Tuple[TN, ...], List[TN]]
TSN = Optional[TS]
TA = Union[T, ARRAY]


class ClipCaptionModel(nn.Module):

    def get_dummy_token(self, batch_size: int, device: torch.device) -> torch.Tensor:
        return torch.zeros(batch_size, self.prefix_length, dtype=torch.int64, device=device)

    def forward(self, tokens: torch.Tensor, prefix: torch.Tensor, mask: Optional[torch.Tensor] = None,
                labels: Optional[torch.Tensor] = None):
        embedding_text = self.gpt.transformer.wte(tokens)
        prefix_projections = self.clip_project(prefix).view(-1, self.prefix_length, self.gpt_embedding_size)
        embedding_cat = torch.cat((prefix_projections, embedding_text), dim=1)
        if labels is not None:
            dummy_token = self.get_dummy_token(tokens.shape[0], tokens.device)
            labels = torch.cat((dummy_token, tokens), dim=1)
        out = self.gpt(inputs_embeds=embedding_cat, labels=labels, attention_mask=mask)
        return out

    def __init__(self):
        super(ClipCaptionModel, self).__init__()
        self.prefix_length = 40
        self.gpt = GPT2LMHeadModel.from_pretrained('gpt2')
        self.gpt_embedding_size = self.gpt.transformer.wte.weight.shape[1]
        self.clip_project = TransformerMapper(640, self.gpt_embedding_size, 40,
                                                                     40, 8)



class MLP(nn.Module):

    def forward(self, x: T) -> T:
        return self.model(x)

    def __init__(self, sizes: Tuple[int, ...], bias=True, act=nn.Tanh):
        super(MLP, self).__init__()
        layers = []
        for i in range(len(sizes) -1):
            layers.append(nn.Linear(sizes[i], sizes[i + 1], bias=bias))
            if i < len(sizes) - 2:
                layers.append(act())
        self.model = nn.Sequential(*layers)


class ClipCaptionPrefix(ClipCaptionModel):

    def parameters(self, recurse: bool = True):
        return self.clip_project.parameters()

    def train(self, mode: bool = True):
        super(ClipCaptionPrefix, self).train(mode)
        self.gpt.eval()
        return self
    
    
class MlpTransformer(nn.Module):
    def __init__(self, in_dim, h_dim, out_d: Optional[int] = None, act=nnf.relu, dropout=0.):
        super().__init__()
        out_d = out_d if out_d is not None else in_dim
        self.fc1 = nn.Linear(in_dim, h_dim)
        self.act = act
        self.fc2 = nn.Linear(h_dim, out_d)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.dropout(x)
        x = self.fc2(x)
        x = self.dropout(x)
        return x


class MultiHeadAttention(nn.Module):

    def __init__(self, dim_self, dim_ref, num_heads, bias=True, dropout=0.):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim_self // num_heads
        self.scale = head_dim ** -0.5
        self.to_queries = nn.Linear(dim_self, dim_self, bias=bias)
        self.to_keys_values = nn.Linear(dim_ref, dim_self * 2, bias=bias)
        self.project = nn.Linear(dim_self, dim_self)
        self.dropout = nn.Dropout(dropout)

    def forward(self, x, y=None, mask=None):
        y = y if y is not None else x
        b, n, c = x.shape
        _, m, d = y.shape
        # b n h dh
        queries = self.to_queries(x).reshape(b, n, self.num_heads, c // self.num_heads)
        # b m 2 h dh
        keys_values = self.to_keys_values(y).reshape(b, m, 2, self.num_heads, c // self.num_heads)
        keys, values = keys_values[:, :, 0], keys_values[:, :, 1]
        attention = torch.einsum('bnhd,bmhd->bnmh', queries, keys) * self.scale
        if mask is not None:
            if mask.dim() == 2:
                mask = mask.unsqueeze(1)
            attention = attention.masked_fill(mask.unsqueeze(3), float("-inf"))
        attention = attention.softmax(dim=2)
        out = torch.einsum('bnmh,bmhd->bnhd', attention, values).reshape(b, n, c)
        out = self.project(out)
        return out, attention


class TransformerLayer(nn.Module):

    def forward_with_attention(self, x, y=None, mask=None):
        x_, attention = self.attn(self.norm1(x), y, mask)
        x = x + x_
        x = x + self.mlp(self.norm2(x))
        return x, attention

    def forward(self, x, y=None, mask=None):
        x = x + self.attn(self.norm1(x), y, mask)[0]
        x = x + self.mlp(self.norm2(x))
        return x

    def __init__(self, dim_self, dim_ref, num_heads, mlp_ratio=4., bias=False, dropout=0., act=nnf.relu,
                 norm_layer: nn.Module = nn.LayerNorm):
        super().__init__()
        self.norm1 = norm_layer(dim_self)
        self.attn = MultiHeadAttention(dim_self, dim_ref, num_heads, bias=bias, dropout=dropout)
        self.norm2 = norm_layer(dim_self)
        self.mlp = MlpTransformer(dim_self, int(dim_self * mlp_ratio), act=act, dropout=dropout)


class Transformer(nn.Module):

    def forward_with_attention(self, x, y=None, mask=None):
        attentions = []
        for layer in self.layers:
            x, att = layer.forward_with_attention(x, y, mask)
            attentions.append(att)
        return x, attentions

    def forward(self, x, y=None, mask=None):
        for i, layer in enumerate(self.layers):
            if i % 2 == 0 and self.enc_dec: # cross
                x = layer(x, y)
            elif self.enc_dec:  # self
                x = layer(x, x, mask)
            else:  # self or cross
                x = layer(x, y, mask)
        return x

    def __init__(self, dim_self: int, num_heads: int, num_layers: int, dim_ref: Optional[int] = None,
                 mlp_ratio: float = 2., act=nnf.relu, norm_layer: nn.Module = nn.LayerNorm, enc_dec: bool = False):
        super(Transformer, self).__init__()
        dim_ref = dim_ref if dim_ref is not None else dim_self
        self.enc_dec = enc_dec
        if enc_dec:
            num_layers = num_layers * 2
        layers = []
        for i in range(num_layers):
            if i % 2 == 0 and enc_dec:  # cross
                layers.append(TransformerLayer(dim_self, dim_ref, num_heads, mlp_ratio, act=act, norm_layer=norm_layer))
            elif enc_dec:  # self
                layers.append(TransformerLayer(dim_self, dim_self, num_heads, mlp_ratio, act=act, norm_layer=norm_layer))
            else:  # self or cross
                layers.append(TransformerLayer(dim_self, dim_ref, num_heads, mlp_ratio, act=act, norm_layer=norm_layer))
        self.layers = nn.ModuleList(layers)
    

class TransformerMapper(nn.Module):

    def forward(self, x):
        x = self.linear(x).view(x.shape[0], self.clip_length, -1)
        prefix = self.prefix_const.unsqueeze(0).expand(x.shape[0], *self.prefix_const.shape)
        prefix = torch.cat((x, prefix), dim=1)
        out = self.transformer(prefix)[:, self.clip_length:]
        return out

    def __init__(self, dim_clip: int, dim_embedding: int, prefix_length: int, clip_length: int, num_layers: int = 8):
        super(TransformerMapper, self).__init__()
        self.clip_length = clip_length
        self.transformer = Transformer(dim_embedding, 8, num_layers)
        self.linear = nn.Linear(dim_clip, clip_length * dim_embedding)
        self.prefix_const = nn.Parameter(torch.randn(prefix_length, dim_embedding), requires_grad=True)