File size: 12,056 Bytes
db6ee6a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#  -------------------------------------------------------------------------------------------
#  Copyright (c) Microsoft Corporation. All rights reserved.
#  Licensed under the MIT License (MIT). See LICENSE in the repo root for license information.
#  -------------------------------------------------------------------------------------------

import math
from dataclasses import dataclass
from functools import partial
from typing import Any, Callable, Optional, Set, Tuple

import torch
import torch.nn as nn
from timm.models.layers import DropPath, Mlp, trunc_normal_


def torch_int_div(tensor1, tensor2):
    """
    A function that performs integer division across different versions of PyTorch.
    """
    return torch.div(tensor1, tensor2, rounding_mode="floor")

@dataclass
class MultiHeadAttentionOutput:
    mha_output: torch.Tensor
    attention: Optional[torch.Tensor] = None


class VisionTransformerPooler(nn.Module):
    """
    :param input_dim: Input feature dimension (i.e., channels in old CNN terminology)
    :param grid_shape: Shape of the grid of patches per image
    :param num_heads: Number of self-attention heads within the MHA block
    :param num_blocks: Number of blocks per attention layer
    :param norm_layer: Normalisation layer

    `self.type_embed`: Is used to characterise prior and current scans, and
                       create permutation variance across modalities/series.
    """

    def __init__(self,
                 input_dim: int,
                 grid_shape: Tuple[int, int],
                 num_heads: int = 8,
                 num_blocks: int = 3,
                 norm_layer: Any = partial(nn.LayerNorm, eps=1e-6)):
        super().__init__()

        block_kwargs = dict(dim=input_dim, num_heads=num_heads, mlp_ratio=1., drop=0.10, attn_drop=0.10,
                            drop_path=0.25, act_layer=nn.GELU, norm_layer=norm_layer)
        self.blocks = nn.ModuleList([Block(**block_kwargs) for _ in range(num_blocks)])
        self.norm_post = norm_layer(input_dim)
        self.grid_shape = grid_shape
        self.num_patches = grid_shape[0] * grid_shape[1]
        self.num_blocks = num_blocks

        # Temporal positional embeddings
        num_series: int = 2
        self.type_embed = nn.Parameter(torch.zeros(num_series, 1, input_dim))
        trunc_normal_(self.type_embed, std=.02)

        # Positional embeddings 1 x L x C (L: Sequence length, C: Feature dimension)
        self.pos_drop = nn.Dropout(p=0.10)
        pos_embed_class = SinePositionEmbedding(embedding_dim=input_dim // 2, normalize=True)
        pos_embed = pos_embed_class(mask=torch.ones([1, grid_shape[0], grid_shape[1]]))  # 1 x L x C
        self.register_buffer("pos_embed", pos_embed, persistent=False)

        # Initialisation
        self.apply(self._init_weights)

    def no_weight_decay(self) -> Set[str]:
        return {'type_embed'}

    def forward(self, current_image: torch.Tensor, previous_image: Optional[torch.Tensor] = None) -> torch.Tensor:
        B, C, H, W = current_image.shape
        assert H == self.grid_shape[0] and W == self.grid_shape[1], "Input and grid shapes do not match"

        # Flatten patch embeddings to have shape (B x L x C), L = H * W
        if previous_image is not None:
            assert previous_image.shape == current_image.shape, "current_image and previous_image shapes do not match"
            previous_image = previous_image.view(B, C, H * W).transpose(1, 2)
        current_image = current_image.view(B, C, H * W).transpose(1, 2)
        pos_embed = self.pos_embed.repeat(B, 1, 1)  # type: ignore

        # Final token activations (B x 2L x C)
        token_features = self.forward_after_reshape(x=current_image, pos_embed=pos_embed, x_previous=previous_image)

        # Extract the patch features of current image
        cur_img_token_id = 0
        current_token_features = token_features[:, cur_img_token_id:self.num_patches+cur_img_token_id]
        current_patch_features = current_token_features.transpose(1, 2).view(B, C, H, W)

        return current_patch_features

    def forward_after_reshape(self,
                              x: torch.Tensor,
                              pos_embed: torch.Tensor,
                              x_previous: Optional[torch.Tensor] = None) -> torch.Tensor:
        B, L, _ = x.shape  # Batch, Sequence length, Feature dimension

        # Positional and type embeddings
        type_embed = self.type_embed[0].expand(B, L, -1)
        if x_previous is not None:
            x = torch.cat((x, x_previous), dim=1)
            pos_embed = torch.cat((pos_embed, pos_embed), dim=1)
            prev_type_embed = self.type_embed[1].expand(B, L, -1)
            type_embed = torch.cat((type_embed, prev_type_embed), dim=1)

        # Add positional and type embeddings (used in query and key matching)
        pos_and_type_embed = pos_embed + type_embed

        # Positional dropout
        x = self.pos_drop(x)

        # Multihead attention followed by MLP
        for block in self.blocks:
            x = block(x=x, pos_and_type_embed=pos_and_type_embed)
        x = self.norm_post(x)

        return x

    def _init_weights(self, m: nn.Module) -> None:
        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)


class MultiHeadAttentionLayer(nn.Module):
    """
    Multi-head self attention module

    The content builds on top of the TIMM library (vision_transformer.py) and differs by the following:
        - Defines a custom `MultiHeadAttentionLayer` which does not only apply `self-attention` but it can be
            generalised to arbitrary (query, key, value) input tuples. This feature can be valuable to process
            more than 2 scans at a time.
        - `Self-attention` specific use-case can still be invoked by calling the `forward_as_mhsa` method.
    """

    def __init__(self,
                 dim: int,
                 num_heads: int = 8,
                 qkv_bias: bool = False,
                 attn_drop: float = 0.,
                 proj_drop: float = 0.) -> None:
        super().__init__()
        self.num_heads = num_heads
        assert dim % num_heads == 0, f"The embedding dim ({dim}) must be divisible by the number of heads ({num_heads})"
        head_dim = dim // num_heads
        self.scale = head_dim ** -0.5
        self.return_attention = False

        self.proj_q = nn.Linear(dim, dim, bias=qkv_bias)
        self.proj_k = nn.Linear(dim, dim, bias=qkv_bias)
        self.proj_v = nn.Linear(dim, dim, bias=qkv_bias)

        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

    def forward(self, k: torch.Tensor, q: torch.Tensor, v: torch.Tensor) -> MultiHeadAttentionOutput:
        B, N, C = v.shape
        assert C % self.num_heads == 0, \
            f"The embedding dim ({C}) must be divisible by the number of heads ({self.num_heads})"

        w_q = self.proj_q(q).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
        w_k = self.proj_k(k).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)
        w_v = self.proj_v(v).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)

        attn = (w_q @ w_k.transpose(-2, -1)) * self.scale
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        o = (attn @ w_v).transpose(1, 2).reshape(B, N, C)
        o = self.proj(o)
        o = self.proj_drop(o)

        attention_output = attn if self.return_attention else None

        return MultiHeadAttentionOutput(mha_output=o, attention=attention_output)

    def forward_as_mhsa(self, input: torch.Tensor) -> MultiHeadAttentionOutput:
        return self(k=input, q=input, v=input)


class Block(nn.Module):
    """
    Encapsulates multi-layer perceptron and multi-head self attention modules into a block.

    The content builds on top of the TIMM library (vision_transformer.py) and differs by the following:
        - This implementation uses spatio-temporal positional embeddings instead of 2D positional embeddings only,
            and they are taken into account within the forward pass of each ViT block.
        - Utilises the custom defined `MultiHeadAttentionLayer` which does not apply `self-attention` only but can be
            generalised to arbitrary (query, key, value) tuples. This can be valuable to process more than 2 scans.

    Positional and type embeddings are handled in a similar fashion as DETR object localisation paper
    https://alcinos.github.io/detr_page/, where a fixed set of sine/cos positional embeddings are used
    in an additive manner to Q and K tensors.
    """

    def __init__(self, dim: int, num_heads: int, mlp_ratio: float = 1., qkv_bias: bool = False, drop: float = 0.,
                 attn_drop: float = 0., drop_path: float = 0., act_layer: Callable = nn.GELU,
                 norm_layer: Callable = nn.LayerNorm) -> None:
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = MultiHeadAttentionLayer(dim=dim, num_heads=num_heads, qkv_bias=qkv_bias,
                                            attn_drop=attn_drop, proj_drop=drop)
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

    def with_pos_and_type_embed(self, tensor: torch.Tensor, emb: Optional[torch.Tensor]) -> torch.Tensor:
        # Add positional embeddings to key and query tensors
        return tensor if emb is None else tensor + emb

    def forward(self, x: torch.Tensor, pos_and_type_embed: Optional[torch.Tensor]) -> torch.Tensor:
        x_with_emb = self.with_pos_and_type_embed(self.norm1(x), emb=pos_and_type_embed)
        x = x + self.drop_path(self.attn.forward_as_mhsa(x_with_emb).mha_output)
        x = x + self.drop_path(self.mlp(self.norm2(x)))

        return x


class SinePositionEmbedding():
    """
    This is a more standard version of the position embedding, very similar to the one used by the Attention is all you
    need paper, generalized to work on images.
    """

    def __init__(self,
                 embedding_dim: int = 64,
                 temperature: int = 10000,
                 normalize: bool = False,
                 scale: float = None) -> None:
        super().__init__()
        self.embedding_dim = embedding_dim
        self.temperature = temperature
        self.normalize = normalize
        if scale is not None and normalize is False:
            raise ValueError("normalize should be True if scale is passed")
        if scale is None:
            scale = 2 * math.pi
        self.scale = scale

    def __call__(self, mask: torch.Tensor) -> torch.Tensor:
        assert mask is not None, "No pixel mask provided"
        B, H, W = mask.shape
        y_embed = mask.cumsum(1, dtype=torch.float32)
        x_embed = mask.cumsum(2, dtype=torch.float32)
        if self.normalize:
            y_embed = y_embed / (y_embed[:, -1:, :] + 1e-6) * self.scale
            x_embed = x_embed / (x_embed[:, :, -1:] + 1e-6) * self.scale

        dim_t = torch.arange(self.embedding_dim, dtype=torch.float32)
        dim_t = self.temperature ** (2 * torch_int_div(dim_t, 2) / self.embedding_dim)

        pos_x = x_embed[:, :, :, None] / dim_t
        pos_y = y_embed[:, :, :, None] / dim_t
        pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
        pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
        pos = torch.cat((pos_y, pos_x), dim=3).view(B, H * W, self.embedding_dim * 2)

        return pos