File size: 16,440 Bytes
9b92c75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
from __future__ import annotations

from copy import deepcopy
from dataclasses import dataclass
from typing import Any

import torch
from torch import Tensor, nn
import torch.nn.functional as F

from .boxes import inverse_sigmoid


class ConvNormAct(nn.Sequential):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        kernel_size: int = 1,
        stride: int = 1,
        groups: int = 1,
        activation: bool = True,
    ) -> None:
        padding = kernel_size // 2
        layers: list[nn.Module] = [
            nn.Conv2d(
                in_channels,
                out_channels,
                kernel_size,
                stride,
                padding,
                groups=groups,
                bias=False,
            ),
            nn.BatchNorm2d(out_channels),
        ]
        if activation:
            layers.append(nn.SiLU(inplace=True))
        super().__init__(*layers)


class GatedConvBlock(nn.Module):
    """Inverted residual block with a cheap learned residual gate."""

    def __init__(self, channels: int, expansion: float = 2.0) -> None:
        super().__init__()
        hidden = int(channels * expansion)
        self.expand = ConvNormAct(channels, hidden)
        self.depthwise = ConvNormAct(hidden, hidden, 3, groups=hidden)
        self.project = ConvNormAct(hidden, channels, activation=False)
        self.gate = nn.Parameter(torch.zeros(1))

    def forward(self, inputs: Tensor) -> Tensor:
        return inputs + torch.tanh(self.gate) * self.project(self.depthwise(self.expand(inputs)))


class BackboneStage(nn.Sequential):
    def __init__(self, in_channels: int, out_channels: int, depth: int, stride: int) -> None:
        super().__init__(
            ConvNormAct(in_channels, out_channels, 3, stride=stride),
            *(GatedConvBlock(out_channels) for _ in range(depth)),
        )


class CompactBackbone(nn.Module):
    def __init__(
        self, stem_channels: int, channels: list[int], depths: list[int]
    ) -> None:
        super().__init__()
        if len(channels) != 4 or len(depths) != 4:
            raise ValueError("Backbone requires four channel and depth values")
        self.stem = nn.Sequential(
            ConvNormAct(3, stem_channels, 3, stride=2),
            ConvNormAct(stem_channels, stem_channels, 3, stride=2),
        )
        stages: list[nn.Module] = []
        in_channels = stem_channels
        for index, (out_channels, depth) in enumerate(zip(channels, depths, strict=True)):
            stages.append(
                BackboneStage(in_channels, out_channels, depth, stride=1 if index == 0 else 2)
            )
            in_channels = out_channels
        self.stages = nn.ModuleList(stages)
        self.out_channels = channels[1:]

    def forward(self, images: Tensor) -> list[Tensor]:
        features = self.stem(images)
        outputs = []
        for index, stage in enumerate(self.stages):
            features = stage(features)
            if index > 0:
                outputs.append(features)
        return outputs


class PyramidFusion(nn.Module):
    def __init__(self, in_channels: list[int], hidden_dim: int, depth: int) -> None:
        super().__init__()
        self.lateral = nn.ModuleList(ConvNormAct(c, hidden_dim) for c in in_channels)
        self.refine = nn.ModuleList(
            nn.Sequential(*(GatedConvBlock(hidden_dim, expansion=1.5) for _ in range(depth)))
            for _ in in_channels
        )

    def forward(self, inputs: list[Tensor]) -> list[Tensor]:
        projected = [layer(x) for layer, x in zip(self.lateral, inputs, strict=True)]
        outputs = list(projected)
        for index in range(len(outputs) - 2, -1, -1):
            outputs[index] = outputs[index] + F.interpolate(
                outputs[index + 1], size=outputs[index].shape[-2:], mode="nearest"
            )
        return [block(x) for block, x in zip(self.refine, outputs, strict=True)]


def sine_position_encoding(
    height: int, width: int, dim: int, device: torch.device, dtype: torch.dtype
) -> Tensor:
    if dim % 4 != 0:
        raise ValueError("Position encoding dimension must be divisible by four")
    y, x = torch.meshgrid(
        torch.linspace(0, 1, height, device=device, dtype=dtype),
        torch.linspace(0, 1, width, device=device, dtype=dtype),
        indexing="ij",
    )
    frequencies = torch.arange(dim // 4, device=device, dtype=dtype)
    frequencies = 2.0 * torch.pi * (10000.0 ** (-frequencies / max(dim // 4, 1)))
    x = x.flatten()[:, None] * frequencies[None]
    y = y.flatten()[:, None] * frequencies[None]
    return torch.cat((x.sin(), x.cos(), y.sin(), y.cos()), dim=-1)


class FeedForward(nn.Sequential):
    def __init__(self, dim: int, expansion: int = 4, dropout: float = 0.0) -> None:
        super().__init__(
            nn.Linear(dim, dim * expansion),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(dim * expansion, dim),
            nn.Dropout(dropout),
        )


class LatentLayer(nn.Module):
    def __init__(self, dim: int, num_heads: int, dropout: float) -> None:
        super().__init__()
        self.norm1 = nn.LayerNorm(dim)
        self.attention = nn.MultiheadAttention(dim, num_heads, dropout, batch_first=True)
        self.norm2 = nn.LayerNorm(dim)
        self.ffn = FeedForward(dim, dropout=dropout)

    def forward(self, inputs: Tensor) -> Tensor:
        normalized = self.norm1(inputs)
        inputs = inputs + self.attention(normalized, normalized, normalized, need_weights=False)[0]
        return inputs + self.ffn(self.norm2(inputs))


class LatentMemory(nn.Module):
    """Compresses multi-scale maps into a fixed-size global reasoning memory."""

    def __init__(
        self,
        dim: int,
        latent_count: int,
        pool_sizes: list[int],
        layers: int,
        num_heads: int,
        dropout: float,
    ) -> None:
        super().__init__()
        if len(pool_sizes) != 3:
            raise ValueError("One latent pool size is required for each pyramid level")
        self.pool_sizes = pool_sizes
        self.latents = nn.Parameter(torch.empty(latent_count, dim))
        self.level_embedding = nn.Parameter(torch.empty(len(pool_sizes), dim))
        self.query_norm = nn.LayerNorm(dim)
        self.token_norm = nn.LayerNorm(dim)
        self.compress = nn.MultiheadAttention(dim, num_heads, dropout, batch_first=True)
        self.layers = nn.ModuleList(LatentLayer(dim, num_heads, dropout) for _ in range(layers))
        nn.init.normal_(self.latents, std=0.02)
        nn.init.normal_(self.level_embedding, std=0.02)

    def forward(self, features: list[Tensor]) -> Tensor:
        tokens = []
        for level, (feature, size) in enumerate(zip(features, self.pool_sizes, strict=True)):
            pooled = F.adaptive_avg_pool2d(feature, (size, size)).flatten(2).transpose(1, 2)
            position = sine_position_encoding(
                size, size, feature.shape[1], feature.device, feature.dtype
            )
            tokens.append(pooled + position[None] + self.level_embedding[level][None, None])
        token_memory = self.token_norm(torch.cat(tokens, dim=1))
        latents = self.latents[None].expand(features[0].shape[0], -1, -1)
        latents = latents + self.compress(
            self.query_norm(latents), token_memory, token_memory, need_weights=False
        )[0]
        for layer in self.layers:
            latents = layer(latents)
        return latents


class QueryLocalSampler(nn.Module):
    """Samples high-resolution pyramid evidence around each evolving query box."""

    def __init__(self, dim: int, num_levels: int, points: int) -> None:
        super().__init__()
        self.num_levels = num_levels
        self.points = points
        self.offsets = nn.Linear(dim, num_levels * points * 2)
        self.weights = nn.Linear(dim, num_levels * points)
        self.output = nn.Linear(dim, dim)
        nn.init.zeros_(self.offsets.weight)
        nn.init.zeros_(self.offsets.bias)
        nn.init.zeros_(self.weights.weight)
        nn.init.zeros_(self.weights.bias)

    def forward(self, queries: Tensor, boxes: Tensor, features: list[Tensor]) -> Tensor:
        batch, query_count, _ = queries.shape
        offsets = self.offsets(queries).view(
            batch, query_count, self.num_levels, self.points, 2
        )
        offsets = offsets.tanh() * boxes[..., None, None, 2:] * 0.5
        centers = boxes[..., None, None, :2]
        sample_points = (centers + offsets).clamp(0.0, 1.0)
        weights = self.weights(queries).view(
            batch, query_count, self.num_levels * self.points
        )
        weights = weights.softmax(dim=-1).view(
            batch, query_count, self.num_levels, self.points
        )

        sampled_levels = []
        for level, feature in enumerate(features):
            grid = sample_points[:, :, level] * 2.0 - 1.0
            sampled = F.grid_sample(
                feature,
                grid,
                mode="bilinear",
                padding_mode="zeros",
                align_corners=False,
            )
            sampled = sampled.permute(0, 2, 3, 1)
            sampled_levels.append(sampled)
        sampled_features = torch.stack(sampled_levels, dim=2)
        fused = (sampled_features * weights[..., None]).sum(dim=(2, 3))
        return self.output(fused)


class DecoderLayer(nn.Module):
    def __init__(
        self, dim: int, num_heads: int, num_levels: int, local_points: int, dropout: float
    ) -> None:
        super().__init__()
        self.norm1 = nn.LayerNorm(dim)
        self.self_attention = nn.MultiheadAttention(dim, num_heads, dropout, batch_first=True)
        self.norm2 = nn.LayerNorm(dim)
        self.global_attention = nn.MultiheadAttention(dim, num_heads, dropout, batch_first=True)
        self.norm3 = nn.LayerNorm(dim)
        self.local_sampler = QueryLocalSampler(dim, num_levels, local_points)
        self.norm4 = nn.LayerNorm(dim)
        self.ffn = FeedForward(dim, dropout=dropout)

    def forward(
        self, queries: Tensor, memory: Tensor, boxes: Tensor, features: list[Tensor]
    ) -> Tensor:
        normalized = self.norm1(queries)
        queries = queries + self.self_attention(
            normalized, normalized, normalized, need_weights=False
        )[0]
        queries = queries + self.global_attention(
            self.norm2(queries), memory, memory, need_weights=False
        )[0]
        queries = queries + self.local_sampler(self.norm3(queries), boxes, features)
        return queries + self.ffn(self.norm4(queries))


class MLP(nn.Sequential):
    def __init__(self, input_dim: int, hidden_dim: int, output_dim: int, layers: int) -> None:
        modules: list[nn.Module] = []
        for index in range(layers):
            in_dim = input_dim if index == 0 else hidden_dim
            out_dim = output_dim if index == layers - 1 else hidden_dim
            modules.append(nn.Linear(in_dim, out_dim))
            if index < layers - 1:
                modules.append(nn.ReLU(inplace=True))
        super().__init__(*modules)


class DenseAuxiliaryHead(nn.Module):
    def __init__(self, dim: int, num_classes: int) -> None:
        super().__init__()
        self.shared = nn.ModuleList(
            nn.Sequential(ConvNormAct(dim, dim, 3, groups=dim), ConvNormAct(dim, dim))
            for _ in range(3)
        )
        self.classification = nn.Conv2d(dim, num_classes, 1)
        self.regression = nn.Conv2d(dim, 4, 1)

    def forward(self, features: list[Tensor]) -> list[dict[str, Tensor]]:
        outputs = []
        for feature, tower in zip(features, self.shared, strict=True):
            hidden = tower(feature)
            outputs.append(
                {
                    "logits": self.classification(hidden),
                    "distances": F.softplus(self.regression(hidden)),
                }
            )
        return outputs


@dataclass(frozen=True)
class ObjectModelV1Spec:
    num_classes: int = 80
    input_size: int = 640
    stem_channels: int = 48
    backbone_channels: tuple[int, int, int, int] = (64, 128, 256, 384)
    backbone_depths: tuple[int, int, int, int] = (2, 3, 6, 3)
    hidden_dim: int = 256
    fpn_depth: int = 2
    latent_count: int = 64
    latent_pool_sizes: tuple[int, int, int] = (12, 6, 3)
    latent_layers: int = 2
    decoder_layers: int = 6
    num_queries: int = 300
    num_heads: int = 8
    local_points: int = 4
    dropout: float = 0.0
    dense_aux: bool = True


class ObjectModelV1(nn.Module):
    """NMS-free detector with compressed global memory and local geometric sampling."""

    def __init__(self, spec: ObjectModelV1Spec) -> None:
        super().__init__()
        self.spec = spec
        self.backbone = CompactBackbone(
            spec.stem_channels, list(spec.backbone_channels), list(spec.backbone_depths)
        )
        self.neck = PyramidFusion(self.backbone.out_channels, spec.hidden_dim, spec.fpn_depth)
        self.memory = LatentMemory(
            spec.hidden_dim,
            spec.latent_count,
            list(spec.latent_pool_sizes),
            spec.latent_layers,
            spec.num_heads,
            spec.dropout,
        )
        decoder_template = DecoderLayer(
            spec.hidden_dim, spec.num_heads, 3, spec.local_points, spec.dropout
        )
        self.decoder = nn.ModuleList(deepcopy(decoder_template) for _ in range(spec.decoder_layers))
        self.query_embedding = nn.Embedding(spec.num_queries, spec.hidden_dim)
        self.reference_points = nn.Embedding(spec.num_queries, 4)
        self.class_heads = nn.ModuleList(
            nn.Linear(spec.hidden_dim, spec.num_classes) for _ in range(spec.decoder_layers)
        )
        self.box_heads = nn.ModuleList(
            MLP(spec.hidden_dim, spec.hidden_dim, 4, 3) for _ in range(spec.decoder_layers)
        )
        self.dense_head = (
            DenseAuxiliaryHead(spec.hidden_dim, spec.num_classes) if spec.dense_aux else None
        )
        self._reset_parameters()

    def _reset_parameters(self) -> None:
        prior_probability = 0.01
        class_bias = -torch.log(torch.tensor((1.0 - prior_probability) / prior_probability))
        for head in self.class_heads:
            nn.init.constant_(head.bias, class_bias)
        nn.init.zeros_(self.reference_points.weight)
        with torch.no_grad():
            self.reference_points.weight[:, 2:] = -2.0
        for head in self.box_heads:
            nn.init.zeros_(head[-1].weight)
            nn.init.zeros_(head[-1].bias)
        if self.dense_head is not None:
            nn.init.constant_(self.dense_head.classification.bias, class_bias)
            nn.init.zeros_(self.dense_head.regression.weight)
            nn.init.constant_(self.dense_head.regression.bias, 1.0)

    def forward(self, images: Tensor) -> dict[str, Any]:
        features = self.neck(self.backbone(images))
        memory = self.memory(features)
        batch = images.shape[0]
        queries = self.query_embedding.weight[None].expand(batch, -1, -1)
        boxes = self.reference_points.weight.sigmoid()[None].expand(batch, -1, -1)
        layer_outputs: list[dict[str, Tensor]] = []
        for layer, class_head, box_head in zip(
            self.decoder, self.class_heads, self.box_heads, strict=True
        ):
            queries = layer(queries, memory, boxes, features)
            boxes = (inverse_sigmoid(boxes) + box_head(queries)).sigmoid()
            layer_outputs.append({"pred_logits": class_head(queries), "pred_boxes": boxes})
            boxes = boxes.detach() if self.training else boxes

        output: dict[str, Any] = dict(layer_outputs[-1])
        output["aux_outputs"] = layer_outputs[:-1]
        if self.training and self.dense_head is not None:
            output["dense_outputs"] = self.dense_head(features)
        return output


def build_model(config: dict[str, Any]) -> ObjectModelV1:
    model_config = config.get("model", config)
    fields = ObjectModelV1Spec.__dataclass_fields__
    unknown = set(model_config) - set(fields)
    if unknown:
        raise ValueError(f"Unknown model configuration keys: {sorted(unknown)}")
    values = dict(model_config)
    for key in ("backbone_channels", "backbone_depths", "latent_pool_sizes"):
        if key in values:
            values[key] = tuple(values[key])
    return ObjectModelV1(ObjectModelV1Spec(**values))