File size: 14,484 Bytes
b73936d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2021 NVIDIA CORPORATION. Licensed under the MIT license.
# Written by Chen Zhu during an internship at NVIDIA, [email protected]
import math

from torch import nn
import torch
from timm.models.layers import trunc_normal_
import torch.nn.functional as F


class AttentionLS(nn.Module):
    """Implementation for long-short term attention.
    Flexible options for using window attention, global token and dynamic projection.

    Args:
        dim: input and output feature dimension.
        num_heads: number of attention heads.
        qkv_bias: whether to use bias for the projection of query, key and values.
        qk_scale: scale factor on query and key for numerical stability.
                  By default, set to square root of head dimensions.
        attn_drop: dropout probability for attention matrix.
        proj_drop: dropout probability for the final output.
        rpe: whether to use relative position encoding.
        nglo: number of global tokens (e.g., CLS).

    """

    def __init__(
        self,
        dim,
        num_heads=8,
        qkv_bias=False,
        qk_scale=None,
        attn_drop=0.0,
        proj_drop=0.0,
        rpe=False,
        nglo=1,
        dp_rank=2,
        w=2,
    ):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
        self.scale = qk_scale or head_dim**-0.5

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)
        self.nglo = nglo

        # Equals to segment size (w) in the paper.
        self.window_size = w
        # Equals to r in the paper.
        self.dp_rank = dp_rank

        if self.dp_rank > 0:
            self.to_dynamic_projection = nn.Linear(dim, dp_rank * num_heads)
        # The LN of DualLN corresponding to dynamic projection
        self.dual_ln_dp = nn.LayerNorm(dim)
        # The LN of DualLN corresponding to all the tokens
        self.dual_ln_full = nn.LayerNorm(dim)

        # Adapted from ViL: https://github.com/microsoft/vision-longformer/blob/main/src/models/layers/longformer2d.py#L55-L100
        # We only add RPE to window attention.
        # Unnecessary to add bias for global tokens, since DualLN already adds biases.
        self.rpe = rpe
        if rpe:
            # handle the boarder conditions...
            w_pad = int(w * 0.5)
            self.local_relative_position_bias_table = nn.Parameter(
                torch.zeros(2 * (w + w_pad - 1) * (2 * w_pad + w + 1) + 1, num_heads)
            )
            trunc_normal_(self.local_relative_position_bias_table, std=0.02)

            # get pair-wise relative position index
            coords_h = torch.arange(-w_pad, w_pad + w)
            coords_w = torch.arange(-w_pad, w_pad + w)
            coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, 2w, 2w
            coords = (
                coords.view(2, (w + w_pad * 2) ** 2).transpose(0, 1).unsqueeze(0)
            )  # 1, 4w**2, 2
            q_coords_hw = torch.arange(0, w)
            q_coords = torch.stack(
                torch.meshgrid([q_coords_hw, q_coords_hw])
            )  # 2, w, w
            q_coords = q_coords.view(2, w**2).transpose(0, 1).unsqueeze(1)  # w**2, 1, 2
            relative_coords = q_coords - coords
            relative_coords += w_pad + w - 1  # shift to start from 0
            relative_coords[:, :, 0] *= 2 * w_pad + w
            relative_position_index = relative_coords.sum(-1)  # w^2, 4w^2
            self.register_buffer("relative_position_index", relative_position_index)

    def forward(self, x, nx=None, ny=None):
        B, N, C = x.shape
        N_feat = N - self.nglo
        self.img_size = int(math.sqrt(N)) if nx is None else nx
        qkv = self.qkv(x)
        # query, key, value
        q, k, v = qkv.chunk(3, dim=2)
        q = q.mul(self.scale)

        # Layer norm on the projected keys and values
        k = self.dual_ln_full(k)
        v = self.dual_ln_full(v)

        # output size: bsz x n_heads x seqlen x d
        if self.nglo > 0:
            q_cls, q = q[:, : self.nglo], q[:, self.nglo :]
            k_cls, k = k[:, : self.nglo], k[:, self.nglo :]
            v_cls, v = v[:, : self.nglo], v[:, self.nglo :]

            q_cls = q_cls.reshape(
                B, self.nglo, self.num_heads, C // self.num_heads
            ).transpose(1, 2)
            k_cls = k_cls.reshape(
                B, self.nglo, self.num_heads, C // self.num_heads
            ).transpose(1, 2)
            v_cls = v_cls.reshape(
                B, self.nglo, self.num_heads, C // self.num_heads
            ).transpose(1, 2)

        q = q.reshape(B, N_feat, self.num_heads, C // self.num_heads).transpose(1, 2)
        k = k.reshape(B, N_feat, self.num_heads, C // self.num_heads).transpose(1, 2)
        v = v.reshape(B, N_feat, self.num_heads, C // self.num_heads).transpose(1, 2)

        # Long-range Attention (Dynamic Projection)
        if self.dp_rank > 0:
            # b x h x r x (l w)
            # Compute the projection matrix (P_i in the paper)
            c_scores = (
                self.to_dynamic_projection(x[:, self.nglo :])
                .transpose(1, 2)
                .contiguous()
                .view(B, self.num_heads, self.dp_rank, -1)
            )
            # c_scores = c_scores.softmax(dim=-1, dtype=torch.float32).to(x)
            c_scores = c_scores.softmax(dim=-1).to(
                x
            )  # Changed when experimenting with mixed precision (Johannes S.)
            # b x h x r x d
            k_lms = c_scores.matmul(k)
            k_lms = k_lms.transpose(1, 2).contiguous().view(B, self.dp_rank, -1)
            k_lms = (
                self.dual_ln_dp(k_lms)
                .view(B, self.dp_rank, self.num_heads, -1)
                .contiguous()
                .permute(0, 2, 3, 1)
            )
            # b x h x (lw) x r
            dots_all = q.matmul(k_lms)

            if self.window_size > 0:
                # Switch the order of dimensions if using window attention.
                dots_all = self.group_dots(dots_all)
        else:
            dots_all = None

        # Short-term Attention (Window Attention)
        # In our window attention, each token attends to at most (4w^2) tokens.
        if self.window_size > 0:
            dots_win = self.compute_window_scores(q, k)
            w2 = int(self.window_size * self.window_size)

            if self.rpe:
                w_pad = int(0.5 * self.window_size)
                local_relative_position_bias = self.local_relative_position_bias_table[
                    self.relative_position_index.view(-1)
                ].view(
                    1, w2, (w_pad * 2 + self.window_size) ** 2, -1
                )  # w^2, kv_nums,H
                local_relative_position_bias = (
                    local_relative_position_bias.permute(0, 3, 1, 2)
                    .expand(B, -1, -1, -1)
                    .unsqueeze(2)
                    .unsqueeze(2)
                )

                dots_win += local_relative_position_bias
            if dots_all is None:
                dots_all = dots_win
            else:
                dots_all = torch.cat([dots_all, dots_win], dim=-1)

        # Global token.
        if self.nglo > 0:
            # and compute the scores of queries on CLS
            dots_q_cls = q.matmul(k_cls.transpose(-1, -2))

            if self.window_size > 0:
                dots_q_cls = self.group_dots(dots_q_cls)
            dots_all = torch.cat([dots_all, dots_q_cls], dim=-1)

        # attn = dots_all.softmax(dim=-1, dtype=torch.float32).to(x)
        attn = dots_all.softmax(dim=-1).to(
            x
        )  # Changed when experimenting with mixed precision (Johannes S.)
        attn = self.attn_drop(attn)
        out = 0
        if self.window_size > 0:
            offset = max(0, self.dp_rank)
            kv_group_size = self.window_size
            total_win_size = max(1, self.window_size // 2) * 2 + kv_group_size
            attn_win = attn[:, :, :, :, :, offset : offset + total_win_size**2]
            out += self.compute_window_pv(attn_win, v)
            attn = self.ungroup_dots(attn)

        # attn will be b x h x lw x n_k from now on
        if self.dp_rank > 0:
            attn_lm = attn[:, :, :, : self.dp_rank]
            v_lms = (
                # c_scores.matmul(v.float())
                c_scores.matmul(
                    v
                )  # Changed when experimenting with mixed precision (Johannes S.)
                .to(v)
                .transpose(1, 2)
                .contiguous()
                .view(B, self.dp_rank, -1)
            )
            v_lms = (
                self.dual_ln_dp(v_lms)
                .view(B, self.dp_rank, self.num_heads, -1)
                .contiguous()
                .transpose(1, 2)
            )

            out += attn_lm.matmul(v_lms)

        if self.nglo > 0:
            attn_cls = attn[:, :, :, -self.nglo :]
            out += attn_cls.matmul(
                v_cls
            )  # Changed. Was `.mul` instead of `.matmul`. (JWS)

            # b x h x 1 x lw
            cls_inner = q_cls.matmul(k_cls.transpose(-1, -2))
            cls_dots = q_cls.matmul(
                k.transpose(-1, -2)
            )  # Changed. Was `out` instead of `k`. (JWS)
            cls_dots = torch.cat([cls_inner, cls_dots], dim=-1)

            # cls_dots = cls_dots.softmax(dim=-1, dtype=torch.float32).to(x)
            cls_dots = cls_dots.softmax(dim=-1).to(
                x
            )  # Changed when experimenting with mixed precision (Johannes S.)
            cls_next = cls_dots[:, :, :, self.nglo :].matmul(
                v
            )  # the post_cls variant # Changed. Was `out` instead of `v`. (JWS)
            cls_next += cls_dots[:, :, :, : self.nglo].matmul(v_cls)

            out = torch.cat([cls_next, out], dim=2)
        out = out.transpose(1, 2).contiguous().view(B, N, -1)

        # x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        out = self.proj(out)
        out = self.proj_drop(out)
        return out

    def compute_window_scores(self, q, k):
        """Compute the inner products for the window attention.
        Frist, divide the query into non-overlapping windows.
        Then, use torch.as_trided (implemented in self.get_overlapping_tiles) to create a view of the keys
        that corresponds to the windows with at most 2x memory overhead.
        Finally, compute the inner product.
        """
        # q: b h (l w) d
        b, h, _, d = q.shape
        side_size = max(self.window_size // 2, 1)
        # q_group_size: segment size
        kv_width = 2 * side_size + self.window_size  # assuming q_stride=1
        q_n_group = self.img_size // self.window_size
        q_tiles = q.reshape(
            b, h, q_n_group, self.window_size, q_n_group, self.window_size, d
        ).permute(0, 1, 2, 4, 3, 5, 6)
        # q_tiles: b x h x n_group x n_group x w^2 x d
        q_tiles = q_tiles.contiguous().view(b, h, q_n_group, q_n_group, -1, d)

        # k_tiles: b x h x n_group x n_group x 9w^2 x d
        k_tiles = (
            self.get_overlapping_tiles(k)
            .contiguous()
            .view(b, h, q_n_group, q_n_group, -1, d)
        )
        # dot_tiles: b x h x n_group x n_group x w^2 x 9w^2
        dot_tiles = q_tiles.matmul(k_tiles.transpose(-1, -2))

        # fill "-inf" into the zero-padding parts
        dot_tiles = dot_tiles.view(b, h, q_n_group, q_n_group, -1, kv_width, kv_width)

        dot_tiles[:, :, 0, :, :, :side_size].fill_(float("-inf"))
        dot_tiles[:, :, -1, :, :, -side_size:].fill_(float("-inf"))
        dot_tiles[:, :, :, 0, :, :, :side_size].fill_(float("-inf"))
        dot_tiles[:, :, :, -1, :, :, -side_size:].fill_(float("-inf"))

        dot_tiles = dot_tiles.view(b, h, q_n_group, q_n_group, -1, kv_width**2)
        return dot_tiles

    def get_overlapping_tiles(self, x):
        """Get overlapping tiles in the 2D spatial domain, ensuring each query computes correlation with all neighbors"""
        # x: b h (l w) d
        b, h, _, d = x.shape
        side_size = max(self.window_size // 2, 1)
        total_size = 2 * side_size + self.window_size
        kv_group_size = self.window_size
        kv_width = self.img_size

        x = x.view(b, h, kv_width, kv_width, d)
        x = F.pad(x, [0, 0, side_size, side_size, side_size, side_size], value=0)

        out_shape = [
            b,
            h,
            kv_width // kv_group_size,
            kv_width // kv_group_size,
            total_size,
            total_size,
            d,
        ]
        in_stride = x.stride()
        out_stride = [
            in_stride[0],
            in_stride[1],
            in_stride[2] * kv_group_size,
            in_stride[3] * kv_group_size,
            in_stride[2],
            in_stride[3],
            in_stride[4],
        ]

        # note we ignored the boundary here
        return x.as_strided(size=out_shape, stride=out_stride)

    def compute_window_pv(self, attn, v):
        """Compute the inner product of attention matrix and the values for the window attention."""
        b, h, n_group, _, w2, n_k = attn.shape
        d = v.shape[-1]
        v_tiles = (
            self.get_overlapping_tiles(v)
            .contiguous()
            .view(b, h, n_group, n_group, -1, d)
        )

        # b x h x n_group x n_group x w^2 x d
        pv = attn.matmul(v_tiles)
        # return: b x h x (lw) x d
        ret = self.ungroup_dots(pv)

        return ret

    def group_dots(self, dots):
        b, h = dots.shape[:2]
        n_group = self.img_size // self.window_size
        dots = dots.reshape(
            b, h, n_group, self.window_size, n_group, self.window_size, -1
        ).permute(0, 1, 2, 4, 3, 5, 6)
        dots = dots.contiguous().view(
            b, h, n_group, n_group, self.window_size * self.window_size, -1
        )
        return dots

    def ungroup_dots(self, dots):
        b, h, n_group, _, _, n_keys = dots.shape
        dots = dots.reshape(
            b, h, n_group, n_group, self.window_size, self.window_size, -1
        ).permute(0, 1, 2, 4, 3, 5, 6)
        dots = dots.contiguous().view(b, h, -1, n_keys)
        return dots