Hugging Face's logo

phanerozoic
/
argus-lite

Image Classification
PyTorch
multi-task-perception
computer-vision
semantic-segmentation
depth-estimation
object-detection
vision-transformer
Model card Files
xet
 Community
phanerozoic commited on
Commit
f65efc0
·
verified ·
1 Parent(s): 30794ab

Argus-Lite: classifier head + full README assembling all four heads

Browse files
Files changed (3) hide show
  1. README.md +88 -0
  2. cls_head.safetensors +3 -0
  3. cls_training_log.json +157 -0
README.md ADDED
@@ -0,0 +1,88 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ license_name: fair-research-license
4
+ license_link: https://huggingface.co/facebook/EUPE-ViT-S/blob/main/LICENSE
5
+ base_model: facebook/EUPE-ViT-S
6
+ tags:
7
+ - multi-task-perception
8
+ - computer-vision
9
+ - image-classification
10
+ - semantic-segmentation
11
+ - depth-estimation
12
+ - object-detection
13
+ - vision-transformer
14
+ library_name: pytorch
15
+ datasets:
16
+ - imagenet-1k
17
+ - scene_parse_150
18
+ - sayakpaul/nyu_depth_v2
19
+ - detection-datasets/coco
20
+ pipeline_tag: image-classification
21
+ ---
22
+
23
+ # Argus-Lite
24
+
25
+ Multi-task perception on a single frozen EUPE-ViT-S backbone, adapted from [phanerozoic/argus](https://huggingface.co/phanerozoic/argus) at roughly ¼ the parameter budget.
26
+
27
+ ## Architecture
28
+
29
+ ```
30
+ Image → EUPE-ViT-S (frozen, 21M) → shared features
31
+ │
32
+ ┌───────────────┼──────────────┬──────────────┐
33
+ ▼ ▼ ▼ ▼
34
+ Classification Segmentation Depth Detection
35
+ Linear(384,1K) BN+Conv(384,150) DPT-style Split-tower (384-D)
36
+ 385 K params 58 K params 1.54 M params 2.91 M params
37
+ ```
38
+
39
+ Plus correspondence via cosine max on patch tokens (0 params).
40
+
41
+ | Component | Params |
42
+ |---|---:|
43
+ | EUPE-ViT-S backbone (frozen) | 21.59 M |
44
+ | Classifier head | 0.39 M |
45
+ | Segmentation head | 0.06 M |
46
+ | Depth head | 1.54 M |
47
+ | Detection head | 2.91 M |
48
+ | **Total** | **~26.5 M** |
49
+
50
+ Roughly **¼** the Argus-B system (103 M) parameter count.
51
+
52
+ ## Training
53
+
54
+ All four heads trained on pre-cached ViT-S features produced by a single forward pass over each target dataset. Backbone is frozen throughout.
55
+
56
+ | Head | Dataset | Input | Recipe | Result |
57
+ |---|---|---|---|---|
58
+ | Classifier | ImageNet-1k train | 224 px CLS token | SGD, lr 30, WD 0, cosine, 30 epochs | 82.87 % train top-1 |
59
+ | Segmentation | ADE20K (20,210 train / 2,000 val) | 512 px, 32×32 grid | AdamW, lr 1e-3, 5 epochs | mIoU 0.417 |
60
+ | Depth | NYUv2 (32K train / 5K val) | 416 px, 26×26 grid, 256-bin | SILog, AdamW, lr 1e-4, 3 epochs | RMSE 0.715 |
61
+ | Detection | COCO train 2017 (117 K) | 768 px, 48×48 grid | FCOS targets, AdamW, lr 1e-4, 2 epochs | loss cls 0.25 / reg 0.33 / ctr 0.60 |
62
+
63
+ Training logs per head live alongside the weights (`*_training_log.json`).
64
+
65
+ ## Files
66
+
67
+ ```
68
+ cls_head.safetensors Linear(384, 1000) classifier
69
+ cls_training_log.json
70
+ seg_head.safetensors BN + Conv2d(384, 150, 1)
71
+ seg_training_log.json
72
+ depth_head.safetensors BN + Conv-Conv-binhead (256 bins)
73
+ depth_training_log.json
74
+ det_head.safetensors SplitTowerHead (feat_dim=384)
75
+ det_training_log.json
76
+ ```
77
+
78
+ ## Caveats
79
+
80
+ These are first-pass trained heads at conservative schedules (one-shot overnight training run). None are at peer-reviewed publication quality:
81
+
82
+ - Classifier top-1 is reported on the training set only (no val cache on disk); run ImageNet val through EUPE-ViT-S + this head to get the public number.
83
+ - Detection head has no mAP evaluation in this release; loss curve converged but COCO-eval requires a separate pass.
84
+ - Depth head is a simpler architecture than Argus-B's DPT; full DPT with hooked intermediate blocks is a natural next iteration.
85
+
86
+ ## Source backbone
87
+
88
+ EUPE-ViT-S from Meta FAIR ([arXiv:2603.22387](https://arxiv.org/abs/2603.22387), Zhu et al., March 2026). Three-stage distillation from PEcore-G + PElang-G + DINOv3-H+ via a 1.9 B proxy teacher. License: FAIR Research License (non-commercial).
cls_head.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a96b191e2f6cf51f35aef4b7d88b0c06e544ff3e2466b855ad317e417e902437
3
+ size 1540152
cls_training_log.json ADDED
@@ -0,0 +1,157 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "head_params": 385000,
3
+ "feat_dim": 384,
4
+ "n_classes": 1000,
5
+ "epochs": [
6
+ {
7
+ "epoch": 1,
8
+ "loss": 1.6970165328146263,
9
+ "train_top1": 0.7001390138834359
10
+ },
11
+ {
12
+ "epoch": 2,
13
+ "loss": 0.846118554977694,
14
+ "train_top1": 0.7827761720369008
15
+ },
16
+ {
17
+ "epoch": 3,
18
+ "loss": 0.7784265583915875,
19
+ "train_top1": 0.7938996243268832
20
+ },
21
+ {
22
+ "epoch": 4,
23
+ "loss": 0.7437803569650671,
24
+ "train_top1": 0.8004225834727244
25
+ },
26
+ {
27
+ "epoch": 5,
28
+ "loss": 0.7212180118059984,
29
+ "train_top1": 0.8049567308555403
30
+ },
31
+ {
32
+ "epoch": 6,
33
+ "loss": 0.7047544569281285,
34
+ "train_top1": 0.8083770499864577
35
+ },
36
+ {
37
+ "epoch": 7,
38
+ "loss": 0.6920474968145377,
39
+ "train_top1": 0.8111799632678643
40
+ },
41
+ {
42
+ "epoch": 8,
43
+ "loss": 0.6813953759093611,
44
+ "train_top1": 0.8135613858302626
45
+ },
46
+ {
47
+ "epoch": 9,
48
+ "loss": 0.6728048469385828,
49
+ "train_top1": 0.815650887042829
50
+ },
51
+ {
52
+ "epoch": 10,
53
+ "loss": 0.6655949958250985,
54
+ "train_top1": 0.8172556739285355
55
+ },
56
+ {
57
+ "epoch": 11,
58
+ "loss": 0.6591138521068679,
59
+ "train_top1": 0.8188596802758735
60
+ },
61
+ {
62
+ "epoch": 12,
63
+ "loss": 0.6535975528028812,
64
+ "train_top1": 0.8202638688008667
65
+ },
66
+ {
67
+ "epoch": 13,
68
+ "loss": 0.648806539474895,
69
+ "train_top1": 0.8212910572938579
70
+ },
71
+ {
72
+ "epoch": 14,
73
+ "loss": 0.6445780284629649,
74
+ "train_top1": 0.8222737550998426
75
+ },
76
+ {
77
+ "epoch": 15,
78
+ "loss": 0.6407953524315197,
79
+ "train_top1": 0.8232010346816613
80
+ },
81
+ {
82
+ "epoch": 16,
83
+ "loss": 0.6374797319105243,
84
+ "train_top1": 0.8241696827970124
85
+ },
86
+ {
87
+ "epoch": 17,
88
+ "loss": 0.6346055535221118,
89
+ "train_top1": 0.8247847470314175
90
+ },
91
+ {
92
+ "epoch": 18,
93
+ "loss": 0.6321213067433529,
94
+ "train_top1": 0.8254201052634044
95
+ },
96
+ {
97
+ "epoch": 19,
98
+ "loss": 0.6298483973607139,
99
+ "train_top1": 0.8259789707352749
100
+ },
101
+ {
102
+ "epoch": 20,
103
+ "loss": 0.6279231848325951,
104
+ "train_top1": 0.8265245670548804
105
+ },
106
+ {
107
+ "epoch": 21,
108
+ "loss": 0.6262850233867244,
109
+ "train_top1": 0.8269062503170937
110
+ },
111
+ {
112
+ "epoch": 22,
113
+ "loss": 0.6248860519371174,
114
+ "train_top1": 0.8273363269581561
115
+ },
116
+ {
117
+ "epoch": 23,
118
+ "loss": 0.6237272049578941,
119
+ "train_top1": 0.8277937224421172
120
+ },
121
+ {
122
+ "epoch": 24,
123
+ "loss": 0.6227332369315073,
124
+ "train_top1": 0.8279076810439232
125
+ },
126
+ {
127
+ "epoch": 25,
128
+ "loss": 0.6219733906739626,
129
+ "train_top1": 0.8280450557967852
130
+ },
131
+ {
132
+ "epoch": 26,
133
+ "loss": 0.621363224941545,
134
+ "train_top1": 0.8283596127593046
135
+ },
136
+ {
137
+ "epoch": 27,
138
+ "loss": 0.6208897446307997,
139
+ "train_top1": 0.8285313312003821
140
+ },
141
+ {
142
+ "epoch": 28,
143
+ "loss": 0.6205545354158298,
144
+ "train_top1": 0.8285695775804404
145
+ },
146
+ {
147
+ "epoch": 29,
148
+ "loss": 0.6203251563298505,
149
+ "train_top1": 0.8286569978777162
150
+ },
151
+ {
152
+ "epoch": 30,
153
+ "loss": 0.620208578519254,
154
+ "train_top1": 0.8286585589544533
155
+ }
156
+ ]
157
+ }