File size: 10,506 Bytes
6e445f1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# --------------------------------------------------------
# Neighborhood Attention Transformer
# Licensed under The MIT License
# Written by Ali Hassani
# --------------------------------------------------------

# Modified by Jitesh Jain

import torch
import torch.nn as nn
from timm.models.layers import DropPath
from detectron2.modeling import BACKBONE_REGISTRY, Backbone, ShapeSpec

from natten import NeighborhoodAttention2D as NeighborhoodAttention


class ConvTokenizer(nn.Module):
    def __init__(self, in_chans=3, embed_dim=96, norm_layer=None):
        super().__init__()
        self.proj = nn.Sequential(
            nn.Conv2d(in_chans, embed_dim // 2, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1)),
            nn.Conv2d(embed_dim // 2, embed_dim, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1)),
        )
        if norm_layer is not None:
            self.norm = norm_layer(embed_dim)
        else:
            self.norm = None

    def forward(self, x):
        x = self.proj(x).permute(0, 2, 3, 1)
        if self.norm is not None:
            x = self.norm(x)
        return x


class ConvDownsampler(nn.Module):
    def __init__(self, dim, norm_layer=nn.LayerNorm):
        super().__init__()
        self.reduction = nn.Conv2d(dim, 2 * dim, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
        self.norm = norm_layer(2 * dim)

    def forward(self, x):
        x = self.reduction(x.permute(0, 3, 1, 2)).permute(0, 2, 3, 1)
        x = self.norm(x)
        return x


class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

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


class NATLayer(nn.Module):
    def __init__(self, dim, num_heads, kernel_size=7, dilation=None,
                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
                 act_layer=nn.GELU, norm_layer=nn.LayerNorm, layer_scale=None):
        super().__init__()
        self.dim = dim
        self.num_heads = num_heads
        self.mlp_ratio = mlp_ratio

        self.norm1 = norm_layer(dim)
        self.attn = NeighborhoodAttention(
            dim, kernel_size=kernel_size, dilation=dilation, num_heads=num_heads,
            qkv_bias=qkv_bias, qk_scale=qk_scale, 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)
        self.mlp = Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio), act_layer=act_layer, drop=drop)
        self.layer_scale = False
        if layer_scale is not None and type(layer_scale) in [int, float]:
            self.layer_scale = True
            self.gamma1 = nn.Parameter(layer_scale * torch.ones(dim), requires_grad=True)
            self.gamma2 = nn.Parameter(layer_scale * torch.ones(dim), requires_grad=True)

    def forward(self, x):
        if not self.layer_scale:
            shortcut = x
            x = self.norm1(x)
            x = self.attn(x)
            x = shortcut + self.drop_path(x)
            x = x + self.drop_path(self.mlp(self.norm2(x)))
            return x
        shortcut = x
        x = self.norm1(x)
        x = self.attn(x)
        x = shortcut + self.drop_path(self.gamma1 * x)
        x = x + self.drop_path(self.gamma2 * self.mlp(self.norm2(x)))
        return x



class NATBlock(nn.Module):
    def __init__(self, dim, depth, num_heads, kernel_size,  dilations=None,
                 downsample=True,
                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., norm_layer=nn.LayerNorm, layer_scale=None):
        super().__init__()
        self.dim = dim
        self.depth = depth

        self.blocks = nn.ModuleList([
            NATLayer(dim=dim,
                     num_heads=num_heads,
                     kernel_size=kernel_size,
                     dilation=None if dilations is None else dilations[i],
                     mlp_ratio=mlp_ratio,
                     qkv_bias=qkv_bias, qk_scale=qk_scale,
                     drop=drop, attn_drop=attn_drop,
                     drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
                     norm_layer=norm_layer,
                     layer_scale=layer_scale)
            for i in range(depth)])

        self.downsample = None if not downsample else ConvDownsampler(dim=dim, norm_layer=norm_layer)

    def forward(self, x):
        for blk in self.blocks:
            x = blk(x)
        if self.downsample is None:
            return x, x
        return self.downsample(x), x


class DiNAT(nn.Module):
    def __init__(self,
                 embed_dim,
                 mlp_ratio,
                 depths,
                 num_heads,
                 drop_path_rate=0.2,
                 in_chans=3,
                 kernel_size=7,
                 dilations=None,
                 out_indices=(0, 1, 2, 3),
                 qkv_bias=True,
                 qk_scale=None,
                 drop_rate=0.,
                 attn_drop_rate=0.,
                 norm_layer=nn.LayerNorm,
                 frozen_stages=-1,
                 layer_scale=None,
                 **kwargs):
        super().__init__()
        self.num_levels = len(depths)
        self.embed_dim = embed_dim
        self.num_features = [int(embed_dim * 2 ** i) for i in range(self.num_levels)]
        self.mlp_ratio = mlp_ratio

        self.patch_embed = ConvTokenizer(in_chans=in_chans, embed_dim=embed_dim, norm_layer=norm_layer)

        self.pos_drop = nn.Dropout(p=drop_rate)

        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]
        self.levels = nn.ModuleList()
        for i in range(self.num_levels):
            level = NATBlock(dim=int(embed_dim * 2 ** i),
                             depth=depths[i],
                             num_heads=num_heads[i],
                             kernel_size=kernel_size,
                             dilations=None if dilations is None else dilations[i],
                             mlp_ratio=self.mlp_ratio,
                             qkv_bias=qkv_bias, qk_scale=qk_scale,
                             drop=drop_rate, attn_drop=attn_drop_rate,
                             drop_path=dpr[sum(depths[:i]):sum(depths[:i + 1])],
                             norm_layer=norm_layer,
                             downsample=(i < self.num_levels - 1),
                             layer_scale=layer_scale)
            self.levels.append(level)

        # add a norm layer for each output
        self.out_indices = out_indices
        for i_layer in self.out_indices:
            layer = norm_layer(self.num_features[i_layer])
            layer_name = f'norm{i_layer}'
            self.add_module(layer_name, layer)

        self.frozen_stages = frozen_stages

    def _freeze_stages(self):
        if self.frozen_stages >= 0:
            self.patch_embed.eval()
            for param in self.patch_embed.parameters():
                param.requires_grad = False

        if self.frozen_stages >= 2:
            for i in range(0, self.frozen_stages - 1):
                m = self.network[i]
                m.eval()
                for param in m.parameters():
                    param.requires_grad = False

    def train(self, mode=True):
        super(DiNAT, self).train(mode)
        self._freeze_stages()

    def forward_embeddings(self, x):
        x = self.patch_embed(x)
        return x

    def forward_tokens(self, x):
        outs = {}
        for idx, level in enumerate(self.levels):
            x, xo = level(x)
            if idx in self.out_indices:
                norm_layer = getattr(self, f'norm{idx}')
                x_out = norm_layer(xo)
                outs["res{}".format(idx + 2)] = x_out.permute(0, 3, 1, 2).contiguous()
        return outs

    def forward(self, x):
        x = self.forward_embeddings(x)
        return self.forward_tokens(x)


@BACKBONE_REGISTRY.register()
class D2DiNAT(DiNAT, Backbone):
    def __init__(self, cfg, input_shape):
        
        embed_dim = cfg.MODEL.DiNAT.EMBED_DIM
        mlp_ratio = cfg.MODEL.DiNAT.MLP_RATIO
        depths = cfg.MODEL.DiNAT.DEPTHS
        num_heads = cfg.MODEL.DiNAT.NUM_HEADS
        drop_path_rate = cfg.MODEL.DiNAT.DROP_PATH_RATE
        kernel_size = cfg.MODEL.DiNAT.KERNEL_SIZE
        out_indices = cfg.MODEL.DiNAT.OUT_INDICES
        dilations = cfg.MODEL.DiNAT.DILATIONS

        super().__init__(
            embed_dim=embed_dim,
            mlp_ratio=mlp_ratio,
            depths=depths,
            num_heads=num_heads,
            drop_path_rate=drop_path_rate,
            kernel_size=kernel_size,
            out_indices=out_indices,
            dilations=dilations,
        )

        self._out_features = cfg.MODEL.DiNAT.OUT_FEATURES

        self._out_feature_strides = {
            "res2": 4,
            "res3": 8,
            "res4": 16,
            "res5": 32,
        }
        self._out_feature_channels = {
            "res2": self.num_features[0],
            "res3": self.num_features[1],
            "res4": self.num_features[2],
            "res5": self.num_features[3],
        }

    def forward(self, x):
        """
        Args:
            x: Tensor of shape (N,C,H,W). H, W must be a multiple of ``self.size_divisibility``.
        Returns:
            dict[str->Tensor]: names and the corresponding features
        """
        assert (
            x.dim() == 4
        ), f"DiNAT takes an input of shape (N, C, H, W). Got {x.shape} instead!"
        outputs = {}
        y = super().forward(x)
        for k in y.keys():
            if k in self._out_features:
                outputs[k] = y[k]
        return outputs

    def output_shape(self):
        return {
            name: ShapeSpec(
                channels=self._out_feature_channels[name], stride=self._out_feature_strides[name]
            )
            for name in self._out_features
        }

    @property
    def size_divisibility(self):
        return 32