File size: 12,915 Bytes
33eda72
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
from collections import OrderedDict
import math
import requests
from io import BytesIO
from functools import partial
from PIL import Image
from typing import Callable, Optional, Sequence, Tuple, List

import torch
from torch import nn
from torch.nn import functional as F
from torch.utils.checkpoint import checkpoint
from torch.nn.init import trunc_normal_
from torchvision import transforms
from torchvision.transforms import InterpolationMode


def get_abs_pos(abs_pos, tgt_size):
    # abs_pos: L, C
    # tgt_size: M
    # return: M, C
    src_size = int(math.sqrt(abs_pos.size(0)))
    tgt_size = int(math.sqrt(tgt_size))
    dtype = abs_pos.dtype

    if src_size != tgt_size:
        return F.interpolate(
            abs_pos.float().reshape(1, src_size, src_size, -1).permute(0, 3, 1, 2),
            size=(tgt_size, tgt_size),
            mode="bicubic",
            align_corners=False,
        ).permute(0, 2, 3, 1).flatten(0, 2).to(dtype=dtype)
    else:
        return abs_pos


class Resampler(nn.Module):
    def __init__(
            self,
            grid_size,
            embed_dim,
            num_heads,
            kv_dim=None,
            norm_layer=nn.LayerNorm
    ):
        super().__init__()
        self.num_queries = grid_size ** 2
        self.embed_dim = embed_dim
        self.num_heads = num_heads

        self.pos_embed = nn.Parameter(torch.randn(embed_dim, grid_size)).requires_grad_(False)

        self.query = nn.Parameter(torch.zeros(self.num_queries, embed_dim))
        trunc_normal_(self.query, std=.02)

        if kv_dim is not None and kv_dim != embed_dim:
            self.kv_proj = nn.Linear(kv_dim, embed_dim, bias=False)
        else:
            self.kv_proj = nn.Identity()

        self.attn = nn.MultiheadAttention(embed_dim, num_heads)
        self.ln_q = norm_layer(embed_dim)
        self.ln_kv = norm_layer(embed_dim)
        
        self.apply(self._init_weights)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)

    def forward(self, x, attn_mask=None):

        pos_embed = get_abs_pos(self.pos_embed, x.size(1))

        x = self.kv_proj(x)
        x = self.ln_kv(x).permute(1, 0, 2)

        N = x.shape[1]
        q = self.ln_q(self.query)
        out = self.attn(
            self._repeat(q, N) + self.pos_embed.unsqueeze(1),
            x + pos_embed.unsqueeze(1),
            x,
            attn_mask=attn_mask)[0]
        return out.permute(1, 0, 2)

    def _repeat(self, query, N: int):
        return query.unsqueeze(1).repeat(1, N, 1)


class VisualAttention(nn.Module):
    """self-attention layer class.

    Self-attention layer takes input with size [s, b, h]
    and returns output of the same size.
    """

    def __init__(self, embed_dim, num_heads,
                 bias=True, kdim=None, vdim=None):
        super(VisualAttention, self).__init__()
        self.embed_dim = embed_dim
        self.kdim = kdim if kdim is not None else embed_dim
        self.vdim = vdim if vdim is not None else embed_dim
        self._qkv_same_embed_dim = self.kdim == embed_dim and self.vdim == embed_dim

        self.num_heads = num_heads

        # Per attention head and per partition values.
        assert embed_dim % num_heads == 0
        self.hidden_size_per_attention_head = embed_dim // num_heads
        self.num_attention_heads_per_partition = num_heads
        self.hidden_size_per_partition = embed_dim

        # Strided linear layer.
        assert self._qkv_same_embed_dim, 'Only Support SelfAttention Currently'
        self.in_proj = nn.Linear(embed_dim, 3 * embed_dim)
        self.out_proj = nn.Linear(embed_dim, embed_dim)
        self.norm_factor = math.sqrt(self.hidden_size_per_attention_head)

    def forward(self, query, key, value, attn_mask = None):
        # query/key/value: [sq, b, h]
        sq, b, _ = query.size()

        assert query is key, 'Only Support Self-Attention Currently'
        sk = sq
        mixed_x_layer = self.in_proj(query)

        # [sq, b, (np * 3 * hn)] --> [sq, b, np, 3 * hn]
        new_tensor_shape = mixed_x_layer.size()[:-1] + \
            (self.num_attention_heads_per_partition,
             3 * self.hidden_size_per_attention_head)
        mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)

        # [sq, b, np, 3 * hn] --> 3 [sq, b, np, hn]
        query_layer, key_layer, value_layer = mixed_x_layer.split(
            self.hidden_size_per_attention_head, dim=-1)

        # [sq, b, np, hn] -> [sq, b * np, hn]
        query_layer = query_layer.view(sq,
            b * self.num_attention_heads_per_partition,
            self.hidden_size_per_attention_head).transpose(0, 1)
        # [sk, b, np, hn] -> [sk, b * np, hn]
        key_layer = key_layer.view(sk,
            b * self.num_attention_heads_per_partition,
            self.hidden_size_per_attention_head).transpose(0, 1)

        q_scaled = query_layer / self.norm_factor
        if attn_mask is not None:
            attention_probs = torch.baddbmm(attn_mask, q_scaled, key_layer.transpose(-2, -1))
        else:
            attention_probs = torch.bmm(q_scaled, key_layer.transpose(-2, -1))
        attention_probs = attention_probs.softmax(dim=-1)

        value_layer = value_layer.view(sk,
            b * self.num_attention_heads_per_partition,
            self.hidden_size_per_attention_head).transpose(0, 1)

        # matmul: [b * np, sq, hn]
        context_layer = torch.bmm(attention_probs, value_layer)

        # change view [b, np, sq, hn]
        context_layer = context_layer.view(b,
            self.num_attention_heads_per_partition,
            sq, self.hidden_size_per_attention_head)

        # [b, np, sq, hn] --> [sq, b, np, hn]
        context_layer = context_layer.permute(2, 0, 1, 3).contiguous()

        # [sq, b, np, hn] --> [sq, b, hp]
        new_context_layer_shape = context_layer.size()[:-2] + \
            (self.hidden_size_per_partition,)
        context_layer = context_layer.view(*new_context_layer_shape)

        output = self.out_proj(context_layer)

        return output


class VisualAttentionBlock(nn.Module):
    def __init__(
            self,
            d_model: int,
            n_head: int,
            mlp_ratio: float = 4.0,
            act_layer: Callable = nn.GELU,
            norm_layer: Callable = nn.LayerNorm,
            is_cross_attention: bool = False,
    ):
        super().__init__()

        self.ln_1 = norm_layer(d_model)
        if is_cross_attention:
            self.ln_1_kv = norm_layer(d_model)

        self.ln_2 = norm_layer(d_model)
        mlp_width = int(d_model * mlp_ratio)
        self.attn = VisualAttention(d_model, n_head)
        self.mlp = nn.Sequential(OrderedDict([
            ("c_fc", nn.Linear(d_model, mlp_width)),
            ("gelu", act_layer()),
            ("c_proj", nn.Linear(mlp_width, d_model))
        ]))

    def attention(
            self,
            q_x: torch.Tensor,
            k_x: Optional[torch.Tensor] = None,
            v_x: Optional[torch.Tensor] = None,
            attn_mask: Optional[torch.Tensor] = None,
    ):
        k_x = k_x if k_x is not None else q_x
        v_x = v_x if v_x is not None else q_x

        attn_mask = attn_mask.to(q_x.dtype) if attn_mask is not None else None
        return self.attn(q_x, k_x, v_x, attn_mask=attn_mask)

    def forward(
            self,
            q_x: torch.Tensor,
            k_x: Optional[torch.Tensor] = None,
            v_x: Optional[torch.Tensor] = None,
            attn_mask: Optional[torch.Tensor] = None,
    ):
        k_x = self.ln_1_kv(k_x) if hasattr(self, "ln_1_kv") and k_x is not None else None
        v_x = self.ln_1_kv(v_x) if hasattr(self, "ln_1_kv") and v_x is not None else None

        x = q_x + self.attention(q_x=self.ln_1(q_x), k_x=k_x, v_x=v_x, attn_mask=attn_mask)
        x = x + self.mlp(self.ln_2(x))
        return x


class Transformer(nn.Module):
    def __init__(
            self,
            width: int,
            layers: int,
            heads: int,
            mlp_ratio: float = 4.0,
            act_layer: Callable = nn.GELU,
            norm_layer: Callable = nn.LayerNorm,
    ):
        super().__init__()
        self.width = width
        self.layers = layers
        self.grad_checkpointing = False

        self.resblocks = nn.ModuleList([
            VisualAttentionBlock(
                width, heads, mlp_ratio, act_layer=act_layer, norm_layer=norm_layer)
            for _ in range(layers)
        ])

    def get_cast_dtype(self) -> torch.dtype:
        return self.resblocks[0].mlp.c_fc.weight.dtype

    def get_cast_device(self) -> torch.device:
        return self.resblocks[0].mlp.c_fc.weight.device

    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
        for r in self.resblocks:
            if self.grad_checkpointing and not torch.jit.is_scripting():
                # TODO: handle kwargs https://github.com/pytorch/pytorch/issues/79887#issuecomment-1161758372
                x = checkpoint(r, x, None, None, attn_mask)
            else:
                x = r(x, attn_mask=attn_mask)
        return x


class VisionTransformer(nn.Module):

    def __init__(
            self,
            image_size: int,
            patch_size: int,
            width: int,
            layers: int,
            heads: int,
            mlp_ratio: float,
            n_queries: int = 256,
            output_dim: int = 512,
            **kwargs
    ):
        super().__init__()
        image_height, image_width = self.image_size = (image_size, image_size)
        patch_height, patch_width = self.patch_size = (patch_size, patch_size)
        self.grid_size = (image_height // patch_height, image_width // patch_width)
        self.output_dim = output_dim

        mean = (0.48145466, 0.4578275, 0.40821073)
        std = (0.26862954, 0.26130258, 0.27577711)
        self.image_transform = transforms.Compose([
            transforms.Resize(
                (image_size, image_size),
                interpolation=InterpolationMode.BICUBIC
            ),
            transforms.ToTensor(),
            transforms.Normalize(mean=mean, std=std),
        ])

        self.conv1 = nn.Conv2d(in_channels=3, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False)

        # class embeddings and positional embeddings
        scale = width ** -0.5
        self.positional_embedding = nn.Parameter(scale * torch.randn(self.grid_size[0] * self.grid_size[1], width))

        norm_layer = partial(nn.LayerNorm, eps=1e-6)
        act_layer = nn.GELU

        self.ln_pre = norm_layer(width)
        self.transformer = Transformer(
            width,
            layers,
            heads,
            mlp_ratio,
            act_layer=act_layer,
            norm_layer=norm_layer,
        )

        self.attn_pool = Resampler(
            grid_size=int(math.sqrt(n_queries)),
            embed_dim=output_dim,
            num_heads=output_dim // 128,
            kv_dim=width,
            norm_layer=norm_layer,
        )
        self.ln_post = norm_layer(output_dim)
        self.proj = nn.Parameter((output_dim** -0.5) * torch.randn(output_dim, output_dim))

    @torch.jit.ignore
    def set_grad_checkpointing(self, enable=True):
        self.transformer.grad_checkpointing = enable

    def forward(self, x: torch.Tensor):
        x = x.to(
            dtype=self.transformer.get_cast_dtype(),
            device=self.transformer.get_cast_device(),
        )
        # to patches
        x = self.conv1(x)  # shape = [*, width, grid, grid]
        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]

        x = x + get_abs_pos(self.positional_embedding, x.size(1))

        x = self.ln_pre(x)

        x = x.permute(1, 0, 2)  # NLD -> LND
        x = self.transformer(x)
        x = x.permute(1, 0, 2)  # LND -> NLD

        if self.attn_pool:
            x = self.attn_pool(x)
        x = self.ln_post(x)
        x = x @ self.proj

        return x

    def encode(self, image_paths: List[str]):
        images = []
        for image_path in image_paths:
            if image_path.startswith("http://") or image_path.startswith("https://"):
                image = Image.open(requests.get(image_path, stream=True).raw)
            elif image_path.startswith("oss://"):
                raise NotImplementedError
            else:
                image = Image.open(image_path)
            image = image.convert("RGB")
            images.append(self.image_transform(image))
        images = torch.stack(images, dim=0)
        return self(images)