File size: 11,232 Bytes
55f3b2e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
958c1e0
 
 
55f3b2e
 
 
958c1e0
55f3b2e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
958c1e0
 
 
55f3b2e
 
 
 
958c1e0
55f3b2e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
958c1e0
 
 
55f3b2e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
958c1e0
 
55f3b2e
 
 
 
 
 
 
958c1e0
 
 
 
 
 
55f3b2e
 
 
 
 
 
 
 
 
 
 
 
958c1e0
 
55f3b2e
 
 
 
 
 
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
"""Repo-level inference dispatcher.

Loads the weights of any stage in this repo and returns a callable person
detector with the shape:

    score: float (+ = person-scene, − = no person)
    present: bool (score > threshold)

Examples:

    # Baseline, via the Argus HF repo for the backbone
    det = PersonDetector.from_stage('stage_0')

    # Tight-FPR variant of the same
    det = PersonDetector.from_stage('stage_0_tight_fpr')

    # Head-pruned backbone
    det = PersonDetector.from_stage('stage_2b')

    # Specialist student (no Argus backbone needed)
    det = PersonDetector.from_stage('stage_4b')

    # Direct-scalar supervision student (same 3.27M as Stage 4)
    det = PersonDetector.from_stage('stage_4c')

    score, present = det.predict('path/to/image.jpg')

Stage 3 (depth reduction) and Stage 5/5b (circuit-level synthesis) are not
loadable at Python level. Stage 3 is an ablation study; Stages 5/5b are
Verilog.
"""
import json, os, sys, io
from pathlib import Path
from typing import Tuple, Union

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from PIL import Image

HERE = Path(__file__).parent
DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
RES = 768
D = 768


def _norm_input(image: Union[str, Path, Image.Image, np.ndarray, torch.Tensor],
                resolution: int = RES) -> torch.Tensor:
    if isinstance(image, (str, Path)):
        img = Image.open(image).convert('RGB')
    elif isinstance(image, Image.Image):
        img = image.convert('RGB')
    elif isinstance(image, np.ndarray):
        img = Image.fromarray(image).convert('RGB')
    elif isinstance(image, torch.Tensor):
        arr = image.cpu().numpy() if image.ndim == 3 else image[0].cpu().numpy()
        if arr.shape[0] == 3:
            arr = arr.transpose(1, 2, 0)
        img = Image.fromarray((arr * 255).astype('uint8')).convert('RGB')
    else:
        raise TypeError(f'unsupported image type: {type(image)}')
    img = img.resize((resolution, resolution), Image.BILINEAR)
    arr = np.asarray(img, dtype=np.uint8).copy()
    x = torch.from_numpy(arr).permute(2, 0, 1).unsqueeze(0).to(DEVICE).float() / 255.0
    mean = torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1).to(DEVICE)
    std = torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1).to(DEVICE)
    return (x - mean) / std


def _load_classifier(path: Path) -> dict:
    with open(path) as f:
        return json.load(f)


class PersonDetector:
    def __init__(self, forward_fn, pos_dims, neg_dims, threshold):
        self._forward = forward_fn
        self._pos = torch.tensor(pos_dims, dtype=torch.long, device=DEVICE)
        self._neg = torch.tensor(neg_dims, dtype=torch.long, device=DEVICE)
        self._thr = float(threshold)

    @torch.inference_mode()
    def predict(self, image) -> Tuple[float, bool]:
        x = _norm_input(image)
        pooled = self._forward(x)        # (D,) float
        score = (pooled[self._pos].sum() - pooled[self._neg].sum()).item()
        return float(score), bool(score > self._thr)

    @classmethod
    def from_stage(cls, stage: str, argus_repo: str = 'phanerozoic/argus',
                   repo_local: Union[str, Path, None] = None):
        """Load one of the stages by name.

        stage ∈ {
            'stage_0', 'stage_0_tight_fpr', 'stage_1',
            'stage_2a',          # heads masked
            'stage_2b',          # backbone structurally pruned
            'stage_4',           # 3.27M student, per-dim MSE
            'stage_4b',          # 15.67M student, cosine on 768-D
            'stage_4c',          # 3.27M student, scalar-MSE
        }

        argus_repo: HF repo for the EUPE-ViT-B backbone. Used by stage_0,
            stage_0_tight_fpr, stage_1, stage_2a. Stage 2b bundles its own
            pruned backbone. Stages 4, 4b, 4c don't use Argus.

        repo_local: local path to this repo (contains stage_*/ directories).
            Defaults to the directory containing this file.
        """
        root = Path(repo_local) if repo_local else HERE

        if stage in ('stage_0', 'stage_1'):
            return cls._build_argus_variant(root / 'stage_0' / 'classifier.json', argus_repo)
        if stage == 'stage_0_tight_fpr':
            return cls._build_argus_variant(root / 'stage_0_tight_fpr' / 'classifier.json', argus_repo)
        if stage in ('stage_2a', 'stage_2'):
            return cls._build_stage2a(root, argus_repo)
        if stage == 'stage_2b':
            return cls._build_stage2b(root)
        if stage == 'stage_4':
            return cls._build_stage4(root, root / 'stage_4' / 'student_final.safetensors',
                                      student_out_dim=40, student_dim=192, student_depth=6, heads=3)
        if stage == 'stage_4b':
            return cls._build_stage4(root, root / 'stage_4b' / 'student_final.safetensors',
                                      student_out_dim=768, student_dim=384, student_depth=8, heads=6)
        if stage == 'stage_4c':
            return cls._build_stage4(root, root / 'stage_4c' / 'student_final.safetensors',
                                      student_out_dim=40, student_dim=192, student_depth=6, heads=3)
        raise ValueError(f'unknown stage: {stage}')

    # --------------- stage-specific builders ---------------

    @classmethod
    def _build_argus_variant(cls, classifier_json, argus_repo):
        from transformers import AutoModel
        model = AutoModel.from_pretrained(argus_repo, trust_remote_code=True).to(DEVICE).eval()
        c = _load_classifier(classifier_json)

        def fwd(x):
            with torch.autocast('cuda' if DEVICE == 'cuda' else 'cpu', dtype=torch.bfloat16):
                out = model.backbone.forward_features(x)
            patches = out['x_norm_patchtokens'].float().squeeze(0)
            ln = F.layer_norm(patches, [D])
            return ln.max(dim=0).values

        return cls(fwd, c['pos_dims'], c['neg_dims'], c['threshold'])

    @classmethod
    def _build_stage2a(cls, root, argus_repo):
        from transformers import AutoModel
        model = AutoModel.from_pretrained(argus_repo, trust_remote_code=True).to(DEVICE).eval()
        c = _load_classifier(root / 'stage_0' / 'classifier.json')
        # Apply head mask from stage_2 head_importance.json (top 10 most prunable)
        with open(root / 'stage_2' / 'head_importance.json') as f:
            imp = json.load(f)
        HEAD_DIM = 64
        with torch.no_grad():
            for (b, h, _drop) in imp['ranked_most_prunable_first'][:10]:
                model.backbone.blocks[b].attn.proj.weight.data[:, h*HEAD_DIM:(h+1)*HEAD_DIM] = 0.0

        def fwd(x):
            with torch.autocast('cuda' if DEVICE == 'cuda' else 'cpu', dtype=torch.bfloat16):
                out = model.backbone.forward_features(x)
            patches = out['x_norm_patchtokens'].float().squeeze(0)
            ln = F.layer_norm(patches, [D])
            return ln.max(dim=0).values

        return cls(fwd, c['pos_dims'], c['neg_dims'], c['threshold'])

    @classmethod
    def _build_stage2b(cls, root):
        sys.path.insert(0, str(root / 'stage_2b'))
        from load_pruned_backbone import load_stage2b_backbone
        backbone = load_stage2b_backbone(
            str(root / 'stage_2b' / 'pruned_state_dict.safetensors'),
            str(root / 'stage_2b' / 'head_config.json'),
        ).to(DEVICE).eval()
        c = _load_classifier(root / 'stage_0' / 'classifier.json')

        def fwd(x):
            with torch.autocast('cuda' if DEVICE == 'cuda' else 'cpu', dtype=torch.bfloat16):
                out = backbone.forward_features(x)
            patches = out['x_norm_patchtokens'].float().squeeze(0)
            ln = F.layer_norm(patches, [D])
            return ln.max(dim=0).values

        return cls(fwd, c['pos_dims'], c['neg_dims'], c['threshold'])

    @classmethod
    def _build_stage4(cls, root, weights_path, student_out_dim, student_dim, student_depth, heads):
        from safetensors.torch import load_file

        class _Block(nn.Module):
            def __init__(self, dim, h, ratio=4.0):
                super().__init__()
                self.norm1 = nn.LayerNorm(dim)
                self.attn = nn.MultiheadAttention(dim, h, batch_first=True)
                self.norm2 = nn.LayerNorm(dim)
                hidden = int(dim * ratio)
                self.mlp = nn.Sequential(nn.Linear(dim, hidden), nn.GELU(), nn.Linear(hidden, dim))

            def forward(self, x):
                h_, _ = self.attn(self.norm1(x), self.norm1(x), self.norm1(x), need_weights=False)
                x = x + h_
                return x + self.mlp(self.norm2(x))

        class _Student(nn.Module):
            def __init__(self, out_dim, dim, depth, h, patch=16, img=RES):
                super().__init__()
                self.patch = nn.Conv2d(3, dim, patch, stride=patch)
                self.pos = nn.Parameter(torch.zeros(1, (img // patch) ** 2, dim))
                self.blocks = nn.ModuleList([_Block(dim, h) for _ in range(depth)])
                self.norm = nn.LayerNorm(dim)
                self.head = nn.Linear(dim, out_dim)

            def forward(self, x):
                t = self.patch(x).flatten(2).transpose(1, 2)
                t = t + self.pos[:, :t.shape[1]]
                for blk in self.blocks:
                    t = blk(t)
                t = self.norm(t)
                return self.head(t.max(dim=1).values)

        student = _Student(student_out_dim, student_dim, student_depth, heads).to(DEVICE).eval()
        student.load_state_dict(load_file(str(weights_path)))

        # Classifier indexing depends on student output layout:
        # - stage_4 / stage_4c: student emits the 40 classifier-relevant dims
        #   directly (pos at [0:20], neg at [20:40]).
        # - stage_4b: student emits a 768-D vector matching teacher layout;
        #   use Stage 0's pos/neg dims directly.
        if student_out_dim == 40:
            pos, neg = list(range(20)), list(range(20, 40))
        else:
            c = _load_classifier(root / 'stage_0' / 'classifier.json')
            pos, neg = c['pos_dims'], c['neg_dims']
        # student_final.safetensors is the peak-F1 epoch (ep3 for Stage 4,
        # ep10 for Stage 4b/4c). Pull that epoch's threshold, not the last.
        with open(Path(weights_path).parent / 'training_log.json') as f:
            log = json.load(f)
        best = max(log['epochs'], key=lambda e: e.get('F1', 0.0))
        thr = best.get('threshold', 0.0)

        def fwd(x):
            with torch.autocast('cuda' if DEVICE == 'cuda' else 'cpu', dtype=torch.bfloat16):
                out = student(x)
            return out.float().squeeze(0)

        return cls(fwd, pos, neg, thr)


if __name__ == '__main__':
    if len(sys.argv) < 3:
        print('usage: python infer.py <stage> <image> [image ...]')
        print('stages: stage_0, stage_0_tight_fpr, stage_1, stage_2a, stage_2b, '
              'stage_4, stage_4b, stage_4c')
        sys.exit(1)
    stage = sys.argv[1]
    det = PersonDetector.from_stage(stage)
    for path in sys.argv[2:]:
        score, present = det.predict(path)
        print(f'{path}  score={score:+.3f}  present={present}')