File size: 9,063 Bytes
7344bef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved

# pyre-unsafe

"""Necks are the interface between a vision backbone and the rest of the detection model"""

from copy import deepcopy
from typing import List, Optional, Tuple

import torch
import torch.nn as nn
from ..model.data_misc import NestedTensor


class Sam3DualViTDetNeck(nn.Module):
    def __init__(
        self,
        trunk: nn.Module,
        position_encoding: nn.Module,
        d_model: int,
        scale_factors=(4.0, 2.0, 1.0, 0.5),
        add_sam2_neck: bool = False,
    ):
        """
        SimpleFPN neck a la ViTDet
        (From detectron2, very lightly adapted)
        It supports a "dual neck" setting, where we have two identical necks (for SAM3 and SAM2), with different weights

        :param trunk: the backbone
        :param position_encoding: the positional encoding to use
        :param d_model: the dimension of the model
        """
        super().__init__()
        self.trunk = trunk
        self.position_encoding = position_encoding
        self.convs = nn.ModuleList()

        self.scale_factors = scale_factors
        use_bias = True
        dim: int = self.trunk.channel_list[-1]

        for _, scale in enumerate(scale_factors):
            current = nn.Sequential()

            if scale == 4.0:
                current.add_module(
                    "dconv_2x2_0",
                    nn.ConvTranspose2d(dim, dim // 2, kernel_size=2, stride=2),
                )
                current.add_module(
                    "gelu",
                    nn.GELU(),
                )
                current.add_module(
                    "dconv_2x2_1",
                    nn.ConvTranspose2d(dim // 2, dim // 4, kernel_size=2, stride=2),
                )
                out_dim = dim // 4
            elif scale == 2.0:
                current.add_module(
                    "dconv_2x2",
                    nn.ConvTranspose2d(dim, dim // 2, kernel_size=2, stride=2),
                )
                out_dim = dim // 2
            elif scale == 1.0:
                out_dim = dim
            elif scale == 0.5:
                current.add_module(
                    "maxpool_2x2",
                    nn.MaxPool2d(kernel_size=2, stride=2),
                )
                out_dim = dim
            else:
                raise NotImplementedError(f"scale_factor={scale} is not supported yet.")

            current.add_module(
                "conv_1x1",
                nn.Conv2d(
                    in_channels=out_dim,
                    out_channels=d_model,
                    kernel_size=1,
                    bias=use_bias,
                ),
            )
            current.add_module(
                "conv_3x3",
                nn.Conv2d(
                    in_channels=d_model,
                    out_channels=d_model,
                    kernel_size=3,
                    padding=1,
                    bias=use_bias,
                ),
            )
            self.convs.append(current)

        self.sam2_convs = None
        if add_sam2_neck:
            # Assumes sam2 neck is just a clone of the original neck
            self.sam2_convs = deepcopy(self.convs)

    def forward(
        self, tensor_list: List[torch.Tensor]
    ) -> Tuple[
        List[torch.Tensor],
        List[torch.Tensor],
        Optional[List[torch.Tensor]],
        Optional[List[torch.Tensor]],
    ]:
        xs = self.trunk(tensor_list)
        sam3_out, sam3_pos = [], []
        sam2_out, sam2_pos = None, None
        if self.sam2_convs is not None:
            sam2_out, sam2_pos = [], []
        x = xs[-1]  # simpleFPN
        for i in range(len(self.convs)):
            sam3_x_out = self.convs[i](x)
            sam3_pos_out = self.position_encoding(sam3_x_out).to(sam3_x_out.dtype)
            sam3_out.append(sam3_x_out)
            sam3_pos.append(sam3_pos_out)

            if self.sam2_convs is not None:
                sam2_x_out = self.sam2_convs[i](x)
                sam2_pos_out = self.position_encoding(sam2_x_out).to(sam2_x_out.dtype)
                sam2_out.append(sam2_x_out)
                sam2_pos.append(sam2_pos_out)
        return sam3_out, sam3_pos, sam2_out, sam2_pos


class Sam3TriViTDetNeck(nn.Module):
    def __init__(
        self,
        trunk: nn.Module,
        position_encoding: nn.Module,
        d_model: int,
        neck_norm=None,
        scale_factors=(4.0, 2.0, 1.0),
    ):
        """
        SimpleFPN neck with three heads (sam3, interactive, propagation).
        """
        super().__init__()
        self.trunk = trunk
        self.position_encoding = position_encoding
        self.convs = nn.ModuleList()

        self.scale_factors = scale_factors
        use_bias = neck_norm is None
        dim = self.trunk.channel_list[-1]

        for _, scale in enumerate(scale_factors):
            current = nn.Sequential()

            if scale == 4.0:
                current.add_module(
                    "dconv_2x2_0",
                    nn.ConvTranspose2d(dim, dim // 2, kernel_size=2, stride=2),
                )
                current.add_module(
                    "gelu",
                    nn.GELU(),
                )
                current.add_module(
                    "dconv_2x2_1",
                    nn.ConvTranspose2d(dim // 2, dim // 4, kernel_size=2, stride=2),
                )
                out_dim = dim // 4
            elif scale == 2.0:
                current.add_module(
                    "dconv_2x2",
                    nn.ConvTranspose2d(dim, dim // 2, kernel_size=2, stride=2),
                )
                out_dim = dim // 2
            elif scale == 1.0:
                out_dim = dim
            elif scale == 0.5:
                current.add_module(
                    "maxpool_2x2",
                    nn.MaxPool2d(kernel_size=2, stride=2),
                )
                out_dim = dim
            else:
                raise NotImplementedError(f"scale_factor={scale} is not supported yet.")

            current.add_module(
                "conv_1x1",
                nn.Conv2d(
                    in_channels=out_dim,
                    out_channels=d_model,
                    kernel_size=1,
                    bias=use_bias,
                ),
            )
            current.add_module(
                "conv_3x3",
                nn.Conv2d(
                    in_channels=d_model,
                    out_channels=d_model,
                    kernel_size=3,
                    padding=1,
                    bias=use_bias,
                ),
            )
            self.convs.append(current)

        # Assumes the new necks are just clones of the original neck
        self.interactive_convs = deepcopy(self.convs)
        self.propagation_convs = deepcopy(self.convs)

    def forward(
        self,
        tensor_list,
        *,
        need_sam3_out: bool = True,
        need_interactive_out: bool = True,
        need_propagation_out: bool = True,
    ):
        xs = self.trunk(tensor_list)
        sam3_out = []
        interactive_out = []
        propagation_out = []

        sam3_pos = []
        interactive_pos = []
        propagation_pos = []
        x = xs[-1]  # simpleFPN
        # OSS trunk returns plain tensors; onevision trunk returns NestedTensors.
        # Use getattr to handle both in a torch.compile-friendly way.
        x_data = getattr(x, "tensors", x)
        x_mask = getattr(x, "mask", None)
        for _, (conv, interactive_conv, propagation_conv) in enumerate(
            zip(self.convs, self.interactive_convs, self.propagation_convs)
        ):
            if need_sam3_out:
                sam3_conv_out = conv(x_data)
                sam3_x_out = NestedTensor(sam3_conv_out, x_mask)
                sam3_out.append(sam3_x_out)
                sam3_pos.append(
                    self.position_encoding(sam3_conv_out).to(sam3_conv_out.dtype)
                )

            if need_interactive_out:
                interactive_conv_out_t = interactive_conv(x_data)
                interactive_conv_out = NestedTensor(interactive_conv_out_t, x_mask)
                interactive_out.append(interactive_conv_out)
                interactive_pos.append(
                    self.position_encoding(interactive_conv_out_t).to(
                        interactive_conv_out_t.dtype
                    )
                )

            if need_propagation_out:
                propagation_conv_out = propagation_conv(x_data)
                propagation_x_out = NestedTensor(propagation_conv_out, x_mask)
                propagation_out.append(propagation_x_out)
                propagation_pos.append(
                    self.position_encoding(propagation_conv_out).to(
                        propagation_conv_out.dtype
                    )
                )

        return (
            sam3_out,
            sam3_pos,
            interactive_out,
            interactive_pos,
            propagation_out,
            propagation_pos,
        )