File size: 8,415 Bytes
bbfa6f6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import numpy as np
from torch import nn
from torch.nn import functional as F
from torch.nn.init import trunc_normal_
import math

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

def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
    """
    grid_size: int of the grid height and width
    return:
    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
    """
    grid_h = np.arange(grid_size, dtype=np.float32)
    grid_w = np.arange(grid_size, dtype=np.float32)
    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
    grid = np.stack(grid, axis=0)

    grid = grid.reshape([2, 1, grid_size, grid_size])
    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
    if cls_token:
        pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
    return pos_embed


def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
    assert embed_dim % 2 == 0

    # use half of dimensions to encode grid_h
    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)

    emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
    return emb


def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
    """
    embed_dim: output dimension for each position
    pos: a list of positions to be encoded: size (M,)
    out: (M, D)
    """
    assert embed_dim % 2 == 0
    omega = np.arange(embed_dim // 2, dtype=np.float32)
    omega /= embed_dim / 2.
    omega = 1. / 10000**omega  # (D/2,)

    pos = pos.reshape(-1)  # (M,)
    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product

    emb_sin = np.sin(out) # (M, D/2)
    emb_cos = np.cos(out) # (M, D/2)

    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
    return emb


class Resampler(nn.Module):
    """
    A 2D perceiver-resampler network with one cross attention layers by
        (grid_size**2) learnable queries and 2d sincos pos_emb
    Outputs:
        A tensor with the shape of (grid_size**2, embed_dim)
    """
    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.from_numpy(get_2d_sincos_pos_embed(embed_dim, grid_size)).float()
        ).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) # batch_first = False
        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).to(dtype=x.dtype)
        out = self.attn(
            self._repeat(q, N) + self.pos_embed.unsqueeze(1).to(dtype=x.dtype),
            x + pos_embed.unsqueeze(1).to(dtype=x.dtype),
            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 TokenCompressor(nn.Module):
    def __init__(
            self,
            num_compressed_token,
            embed_dim,
            num_heads,
            kv_dim=None,
            norm_layer=nn.LayerNorm
    ):
        super().__init__()
        self.embed_dim = embed_dim
        self.num_heads = num_heads
        self.num_compressed_token = num_compressed_token

        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):
        # zero initializatoin ,identical
        if isinstance(m, nn.Linear):
            # trunc_normal_(m.weight, std=.02)
            nn.init.constant_(m.weight, 0.0)
            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, compress_version='v0'):
        
        topk_index_sorted = self.token_filter(x, compress_version)
        out = torch.gather(x, 1, topk_index_sorted.unsqueeze(-1).expand(-1, -1, x.shape[-1])).permute(1, 0, 2)
        q = torch.gather(x, 1, topk_index_sorted.unsqueeze(-1).expand(-1, -1, x.shape[-1])).clone().detach()
        q = q.permute(1, 0, 2)
        x = self.kv_proj(x)
        x = self.ln_kv(x).permute(1, 0, 2)

        q = self.ln_q(q)
        out += self.attn(
            q,
            x,
            x,
            attn_mask=attn_mask)[0]
        return out.permute(1, 0, 2)
    

    def token_filter(self, x, compress_version='v0'):
        tokens = x.clone().detach()
        tokens_norm = tokens / tokens.norm(dim=-1, keepdim=True)
        attn_scores = torch.matmul(tokens_norm, tokens_norm.transpose(2, 1))
        # mask = torch.ones((tokens_norm.shape[1], tokens_norm.shape[1]), device=tokens_norm.device).triu()
        if compress_version == 'v0':
            mask = torch.eye(tokens_norm.shape[1], device=tokens_norm.device)
        elif compress_version == 'v1':
            mask = torch.ones((tokens_norm.shape[1], tokens_norm.shape[1]), device=tokens_norm.device).triu()
        else:
            NotImplementedError
        attn_scores = attn_scores.masked_fill(mask == 1, 1e-5)
        importances = 1 - attn_scores.max(dim=-1)[0]
        topk_index = torch.topk(importances, self.num_compressed_token)[1]
        topk_index_sorted =  torch.sort(topk_index, dim=-1)[0]
        return topk_index_sorted
    







# from flash_perceiver import Perceiver, utils
# from torchstat import stat
# batch_size, seq_len, in_dim = 32, 5120, 1024

# latent_dim = 1024
# num_latents = 128
# out_dim = 1024

# model = Perceiver(
#     input_dim=in_dim,
#     depth=4,
#     output_dim=out_dim,
#     num_latents=num_latents,
#     latent_dim=latent_dim,
#     cross_heads=1,
#     cross_head_dim=64,
#     cross_rotary_emb_dim=0,
#     cross_attn_dropout=0.0,
#     latent_heads=8,
#     latent_head_dim=64,
#     latent_rotary_emb_dim=0,
#     latent_attn_dropout=0.0,
#     weight_tie_layers=False,
#     gated_mlp=True,
#     self_per_cross_attn=1,
#     num_zero_tokens=None,
#     use_flash_attn=True,
# ).cuda()

# data = torch.randn(batch_size, seq_len, in_dim, device='cuda:0')

# # `out_dim` specified; averages and projects output
# # Note: FlashAttention only supports half-precision.
# #  We need to use `torch.autocast` for the forward-pass
# with torch.autocast('cuda'):
#     out = model(data, return_embeddings=True)
#     print(torch.cuda.max_memory_allocated(device=None))
# print(out.shape)