File size: 12,686 Bytes
7766391
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""NAPGuard Patch Detector model for outpost deployment.

Accepts PIL images directly, applies NFSI preprocessing, runs YOLOv5s
detection, and returns patch detection results.

Usage (inside outpost):
    result = model.predict(image=pil_image)
    # returns {"score": 0.85, "num_detections": 2}

Reference: Wu et al., CVPR 2024
"""

from __future__ import annotations

import sys
from typing import List, Optional

import torch
import torch.nn as nn
import torch.nn.functional as F
from PIL import Image
from torchvision import transforms
from torchvision.ops import nms
from transformers import PreTrainedModel

from .configuration_napguard import NAPGuardPatchDetectorConfig


def _log(msg):
    print(f"[NAPGUARD-DEBUG] {msg}", file=sys.stderr, flush=True)


# ---------------------------------------------------------------------------
# YOLOv5s building blocks
# ---------------------------------------------------------------------------

def _autopad(k, p=None, d=1):
    if d > 1:
        k = d * (k - 1) + 1 if isinstance(k, int) else [d * (x - 1) + 1 for x in k]
    if p is None:
        p = k // 2 if isinstance(k, int) else [x // 2 for x in k]
    return p


class Conv(nn.Module):
    default_act = nn.SiLU()

    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, d=1, act=True):
        super().__init__()
        self.conv = nn.Conv2d(c1, c2, k, s, _autopad(k, p, d), groups=g, dilation=d, bias=False)
        self.bn = nn.BatchNorm2d(c2)
        self.act = self.default_act if act is True else act if isinstance(act, nn.Module) else nn.Identity()

    def forward(self, x):
        return self.act(self.bn(self.conv(x)))


class Bottleneck(nn.Module):
    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5):
        super().__init__()
        c_ = int(c2 * e)
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c_, c2, 3, 1, g=g)
        self.add = shortcut and c1 == c2

    def forward(self, x):
        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))


class C3(nn.Module):
    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
        super().__init__()
        c_ = int(c2 * e)
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c1, c_, 1, 1)
        self.cv3 = Conv(2 * c_, c2, 1)
        self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))

    def forward(self, x):
        return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), 1))


class SPPF(nn.Module):
    def __init__(self, c1, c2, k=5):
        super().__init__()
        c_ = c1 // 2
        self.cv1 = Conv(c1, c_, 1, 1)
        self.cv2 = Conv(c_ * 4, c2, 1, 1)
        self.m = nn.MaxPool2d(kernel_size=k, stride=1, padding=k // 2)

    def forward(self, x):
        x = self.cv1(x)
        y1 = self.m(x)
        y2 = self.m(y1)
        return self.cv2(torch.cat((x, y1, y2, self.m(y2)), 1))


class Detect(nn.Module):
    stride = None

    def __init__(self, nc=1, anchors=(), ch=()):
        super().__init__()
        self.nc = nc
        self.no = nc + 5
        self.nl = len(anchors)
        self.na = len(anchors[0]) // 2
        self.grid = [torch.empty(0) for _ in range(self.nl)]
        self.anchor_grid = [torch.empty(0) for _ in range(self.nl)]
        self.register_buffer('anchors', torch.tensor(anchors).float().view(self.nl, -1, 2))
        self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch)

    def forward(self, x):
        z = []
        for i in range(self.nl):
            x[i] = self.m[i](x[i])
            bs, _, ny, nx = x[i].shape
            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
            if self.grid[i].shape[2:4] != x[i].shape[2:4]:
                self.grid[i], self.anchor_grid[i] = self._make_grid(nx, ny, i)
            xy, wh, conf = x[i].sigmoid().split((2, 2, self.nc + 1), 4)
            xy = (xy * 2 - 0.5 + self.grid[i]) * self.stride[i]
            wh = (wh * 2) ** 2 * self.anchor_grid[i]
            z.append(torch.cat((xy, wh, conf), 4).view(bs, self.na * nx * ny, self.no))
        return (torch.cat(z, 1),)

    def _make_grid(self, nx, ny, i):
        d = self.anchors[i].device
        t = self.anchors[i].dtype
        shape = 1, self.na, ny, nx, 2
        y, x = torch.arange(ny, device=d, dtype=t), torch.arange(nx, device=d, dtype=t)
        yv, xv = torch.meshgrid(y, x, indexing='ij')
        grid = torch.stack((xv, yv), 2).expand(shape) - 0.5
        anchor_grid = (self.anchors[i] * self.stride[i]).view((1, self.na, 1, 1, 2)).expand(shape)
        return grid, anchor_grid


class _Upsample(nn.Module):
    """Placeholder for upsample layers (no parameters, needed for indexing)."""
    def __init__(self):
        super().__init__()
        self.up = nn.Upsample(None, 2, 'nearest')
    def forward(self, x):
        return self.up(x)


class _Concat(nn.Module):
    """Placeholder for concat layers (no parameters, needed for indexing)."""
    def forward(self, x):
        return torch.cat(x, 1)


# ---------------------------------------------------------------------------
# NFSI
# ---------------------------------------------------------------------------

def _nfsi(imgs, sigma=3.0, threshold_factor=2.0):
    # FFT requires float32 (cuFFT doesn't support fp16 for non-power-of-2 sizes)
    orig_dtype = imgs.dtype
    imgs_f32 = imgs.float()

    blur = transforms.GaussianBlur(3, sigma)
    _, _, height, width = imgs_f32.shape
    R = (height + width) // 8
    yy, xx = torch.meshgrid(torch.arange(height), torch.arange(width), indexing="ij")
    lpf = (((xx - (width - 1) / 2) ** 2 + (yy - (height - 1) / 2) ** 2) < R ** 2).float().to(imgs_f32.device)
    im_copy = imgs_f32.clone()
    mask_bg = torch.ones_like(imgs_f32)
    f = torch.fft.fftn(im_copy, dim=(2, 3))
    f = torch.roll(f, (height // 2, width // 2), dims=(2, 3))
    f_l = torch.roll(f * lpf, (-height // 2, -width // 2), dims=(2, 3))
    x_l = torch.abs(torch.fft.ifftn(f_l, dim=(2, 3))).clamp(0, 1).mean(dim=1)
    mu, std = x_l.mean(dim=(1, 2)), x_l.std(dim=(1, 2))
    for idx in range(x_l.shape[0]):
        x_l[idx] = torch.where(
            torch.abs(x_l[idx] - mu[idx]) > threshold_factor * std[idx],
            torch.ones_like(x_l[idx]), torch.zeros_like(x_l[idx]))
    mask = x_l.unsqueeze(1).repeat(1, 3, 1, 1)
    result = blur(im_copy).clamp_(0, 1) * mask + imgs_f32 * (mask_bg - mask)
    return result.to(orig_dtype)


# ---------------------------------------------------------------------------
# HuggingFace wrapper
# ---------------------------------------------------------------------------

# YOLOv5s layer structure (matching original state_dict keys model.0 - model.24)
# Layers 11, 12, 15, 16, 19, 22 are Upsample/Concat (no params)
_ANCHORS = [[10, 13, 16, 30, 33, 23], [30, 61, 62, 45, 59, 119], [116, 90, 156, 198, 373, 326]]


class NAPGuardPatchDetectorModel(PreTrainedModel):
    """NAPGuard YOLOv5s patch detector with state_dict key prefix `model.N.*`.

    The nn.ModuleList index matches the original YOLOv5 layer numbering
    so that state_dict keys align for from_pretrained loading.
    """

    config_class = NAPGuardPatchDetectorConfig
    supports_gradient_checkpointing = False

    def __init__(self, config: NAPGuardPatchDetectorConfig) -> None:
        super().__init__(config)
        self._input_size = config.input_size
        self._conf_thres = config.conf_thres
        self._iou_thres = config.iou_thres
        self._use_nfsi = config.use_nfsi
        self._nfsi_sigma = config.nfsi_sigma
        self._nfsi_threshold_factor = config.nfsi_threshold_factor
        self._to_tensor = transforms.ToTensor()

        # Build model as ModuleList to match state_dict key prefix model.N
        self.model = nn.ModuleList([
            Conv(3, 32, 6, 2, 2),         # 0: P1/2
            Conv(32, 64, 3, 2),            # 1: P2/4
            C3(64, 64, 1),                 # 2
            Conv(64, 128, 3, 2),           # 3: P3/8
            C3(128, 128, 2),               # 4
            Conv(128, 256, 3, 2),          # 5: P4/16
            C3(256, 256, 3),               # 6
            Conv(256, 512, 3, 2),          # 7: P5/32
            C3(512, 512, 1),               # 8
            SPPF(512, 512, 5),             # 9
            Conv(512, 256, 1, 1),          # 10
            _Upsample(),                   # 11 (no params)
            _Concat(),                     # 12 (no params)
            C3(512, 256, 1, False),        # 13
            Conv(256, 128, 1, 1),          # 14
            _Upsample(),                   # 15 (no params)
            _Concat(),                     # 16 (no params)
            C3(256, 128, 1, False),        # 17
            Conv(128, 128, 3, 2),          # 18
            _Concat(),                     # 19 (no params)
            C3(256, 256, 1, False),        # 20
            Conv(256, 256, 3, 2),          # 21
            _Concat(),                     # 22 (no params)
            C3(512, 512, 1, False),        # 23
            Detect(1, _ANCHORS, (128, 256, 512)),  # 24
        ])

        # Set detect stride
        self.model[24].stride = torch.tensor([8., 16., 32.])

        # Save/restore indices for the PANet forward pass
        self._save_indices = {4, 6, 9, 10, 13, 14, 17, 20}

    def forward(self, pixel_values=None, **kwargs):
        if "image" in kwargs:
            return self.predict(**kwargs)
        if pixel_values is None:
            raise ValueError("Provide pixel_values or image=PIL")
        return self._yolo_forward(pixel_values)

    def _yolo_forward(self, x):
        """YOLOv5s forward with PANet skip connections."""
        saved = {}

        # Backbone: layers 0-9
        for i in range(10):
            x = self.model[i](x)
            if i in self._save_indices:
                saved[i] = x

        # Neck: 10 β†’ upsample β†’ cat(P4) β†’ 13
        x = self.model[10](x)
        saved[10] = x
        x = self.model[11](x)                          # upsample
        x = self.model[12]([x, saved[6]])               # cat with P4
        x = self.model[13](x)
        saved[13] = x

        # 14 β†’ upsample β†’ cat(P3) β†’ 17
        x = self.model[14](x)
        saved[14] = x
        x = self.model[15](x)                          # upsample
        x = self.model[16]([x, saved[4]])               # cat with P3
        x = self.model[17](x)
        det_small = x                                    # P3 output
        saved[17] = x

        # 18 β†’ cat(layer 14 output) β†’ 20
        x = self.model[18](x)
        x = self.model[19]([x, saved[14]])               # cat with layer 14
        x = self.model[20](x)
        det_mid = x                                      # P4 output
        saved[20] = x

        # 21 β†’ cat(10 output) β†’ 23
        x = self.model[21](x)
        x = self.model[22]([x, saved[10]])               # cat
        x = self.model[23](x)
        det_large = x                                    # P5 output

        # Detect
        return self.model[24]([det_small, det_mid, det_large])

    @torch.no_grad()
    def predict(self, image: Image.Image, **kwargs) -> dict:
        img = image.convert("RGB").resize(
            (self._input_size, self._input_size), Image.Resampling.BILINEAR
        )
        tensor = self._to_tensor(img).unsqueeze(0).to(device=self.device, dtype=self.dtype)

        if self._use_nfsi:
            tensor = _nfsi(tensor, self._nfsi_sigma, self._nfsi_threshold_factor)

        pred = self._yolo_forward(tensor)
        prediction = pred[0] if isinstance(pred, tuple) else pred
        if prediction.ndim == 3:
            prediction = prediction[0]

        # Filter by obj conf
        xc = prediction[..., 4] > self._conf_thres
        x = prediction[xc]
        if x.shape[0] == 0:
            return {"score": 0.0, "num_detections": 0}

        # Combine obj conf * class conf
        x[:, 5:] *= x[:, 4:5]
        conf, cls = x[:, 5:].max(1, keepdim=True)
        x = torch.cat([x[:, :4], conf, cls], dim=1)
        x = x[x[:, 4] > self._conf_thres]
        if x.shape[0] == 0:
            return {"score": 0.0, "num_detections": 0}

        # xywh β†’ xyxy
        boxes = x[:, :4].clone()
        boxes[:, 0] = x[:, 0] - x[:, 2] / 2
        boxes[:, 1] = x[:, 1] - x[:, 3] / 2
        boxes[:, 2] = x[:, 0] + x[:, 2] / 2
        boxes[:, 3] = x[:, 1] + x[:, 3] / 2

        keep = nms(boxes, x[:, 4], self._iou_thres)
        x = x[keep]

        return {"score": float(x[:, 4].max().item()), "num_detections": int(x.shape[0])}

    @torch.no_grad()
    def score_image(self, image: Image.Image, **kwargs) -> dict:
        return self.predict(image=image, **kwargs)