adithya-balaji commited on
Commit
7307975
·
verified ·
1 Parent(s): 97719eb

Add/update model card for medium

Browse files
Files changed (1) hide show
  1. README.md +98 -0
README.md ADDED
@@ -0,0 +1,98 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ library_name: rfdetr.cpp
4
+ tags:
5
+ - object-detection
6
+ - rfdetr
7
+ - gguf
8
+ - ggml
9
+ - cpp-inference
10
+ pipeline_tag: object-detection
11
+ base_model: roboflow/rfdetr
12
+ ---
13
+
14
+ # RF-DETR Medium — GGUF for rfdetr.cpp
15
+
16
+ GGUF-format weights of [Roboflow RF-DETR Medium](https://github.com/roboflow/rf-detr) (detection variant) for use with [rfdetr.cpp](https://github.com/adithyab94/rf-detr.cpp), a C++/ggml implementation that matches the upstream PyTorch model on CPU.
17
+
18
+ This repo contains all four standard quantizations of this variant. **F16 is the recommended default** — same accuracy as F32, 1.85× smaller, and typically the fastest on modern CPUs thanks to ggml's F32×F16 matmul fast path.
19
+
20
+ ## Available files
21
+
22
+ | File | Quant | Size (MB) | Recall @ IoU 0.5 | Recall @ IoU 0.95 | Mean \|Δscore\| | Latency (median ms, T=8) |
23
+ |---|---|---:|---:|---:|---:|---:|
24
+ | `rfdetr-medium-f32.gguf` | F32 | 125.0 | 1.0000 | 1.0000 | 0.0008 | 139.9 |
25
+ | `rfdetr-medium-f16.gguf` ← **recommended** | F16 | 67.2 | 1.0000 | 1.0000 | 0.0023 | 143.9 |
26
+ | `rfdetr-medium-q8_0.gguf` | Q8_0 | 40.2 | 1.0000 | 1.0000 | 0.0063 | 145.8 |
27
+ | `rfdetr-medium-q4_K.gguf` | Q4_K | 32.5 | 1.0000 | 1.0000 | 0.0230 | 155.8 |
28
+
29
+ All accuracy numbers above are computed against the upstream PyTorch reference (`rfdetr 1.9.0`) on 1 image (test_image.jpg) at threshold 0.5. Latency is measured with `rfdetr-cli bench` (8 iters + 3 warmup) at T=8 threads on a single Intel Core i7-12800HX image (test_image.jpg).
30
+
31
+ ## Architecture
32
+
33
+ - Backbone: DINOv2-small
34
+ - Input resolution: 576×576
35
+ - Patch size: 16
36
+ - Decoder layers: 4
37
+ - Object queries: 300
38
+ - Task: object detection (boxes only)
39
+
40
+ ## Quantization notes
41
+
42
+ - **F32** — full-precision reference, ~120 MB. Bit-exact PyTorch parity.
43
+ - **F16** — matmul-multiplicand weights only; LayerNorms, conv kernels, embeddings, biases, and layer-scale gammas stay F32. Lossless on this model and consistently the fastest variant on CPU.
44
+ - **Q8_0** — best size/accuracy tradeoff under F16; ~3× smaller than F32 with effectively identical detections.
45
+ - **Q4_K** — smallest practical quant. Rows with `ne[0] % 256 != 0` (the decoder's 128-dim MLP halves, 60 tensors) silently fall back to Q8_0 per ggml's quantizer logic — net compression is still ~3.8× over F32. Use only when the size budget is tight; expect a measurable Recall@0.95 drop relative to F16/Q8_0 (see file table above).
46
+
47
+ ## Compatibility
48
+
49
+ These GGUFs stamp `rfdetr.preprocess.resize_mode = "bilinear_no_antialias"`, matching RF-DETR 1.9's antialias-free float bilinear resize (`align_corners=false`, half-pixel coordinates, no intermediate uint8 rounding). rf-detr.cpp treats this key as **optional**: GGUFs that predate this metadata (no `resize_mode` key) keep using the legacy stb-based resize path, so older files continue to produce their original outputs unchanged. An unrecognized `resize_mode` value is rejected rather than guessed.
50
+
51
+ **Keypoint-preview inference is not supported.** rf-detr.cpp does not implement the keypoint output head; this repository only serves box detection outputs.
52
+
53
+ ## Usage
54
+
55
+ ```bash
56
+ # 1. Clone + build rfdetr.cpp
57
+ git clone https://github.com/adithyab94/rf-detr.cpp
58
+ cd rf-detr.cpp
59
+ cmake -B build -DRFDETR_BUILD_CLI=ON && cmake --build build -j
60
+
61
+ # 2. Download a quant (F16 recommended)
62
+ hf download adithya-balaji/rfdetr-cpp-medium rfdetr-medium-f16.gguf --local-dir models/
63
+
64
+ # 3. Run detection
65
+ build/bin/rfdetr-cli detect \
66
+ --model models/rfdetr-medium-f16.gguf \
67
+ --input my_image.jpg \
68
+ --threshold 0.5 --threads 8 \
69
+ --output detections.json
70
+ ```
71
+
72
+ ## Accuracy methodology
73
+
74
+ All accuracy metrics are computed against the upstream PyTorch reference (`rfdetr 1.9.0`) on 1 image (test_image.jpg) at threshold 0.5. Each detection match uses greedy Hungarian-style assignment by IoU (≥ 0.5 lenient, ≥ 0.95 strict) with class equality required.
75
+
76
+ See [BENCHMARK.md](https://github.com/adithyab94/rf-detr.cpp/blob/main/BENCHMARK.md) and [`benchmarks/results/accuracy_sweep.json`](https://github.com/adithyab94/rf-detr.cpp/blob/main/benchmarks/results/accuracy_sweep.json) for the full sweep across the (variant × quant) cells.
77
+
78
+ ## Provenance
79
+
80
+ - Source project: [Roboflow RF-DETR](https://github.com/roboflow/rf-detr)
81
+ - Upstream package: `rfdetr==1.9.0`
82
+ - Converted with [rfdetr.cpp](https://github.com/adithyab94/rf-detr.cpp) at commit [`f0fd7f4ff643`](https://github.com/adithyab94/rf-detr.cpp/commit/f0fd7f4ff643824862071396cafeaf691f3cc800)
83
+ - Checkpoint: official pretrained `rfdetr-medium` weights (downloaded by the `rfdetr` package on first use)
84
+
85
+ ### Checksums (SHA-256)
86
+
87
+ Also available as `SHA256SUMS` in this repo.
88
+
89
+ ```
90
+ 664ffe389100cbc942a5d99a2d5b48ef12fdf88f047471fd42d98c5c7d8d32fb rfdetr-medium-f32.gguf
91
+ 353bdccb3a6aa4865fe5c6b636d627c7d6f8c6f94701b9d074d5c2b757b4c792 rfdetr-medium-f16.gguf
92
+ b8f77932ceb8429856588aabfa415a5a752eba2d6c212b043faf473a8aab575f rfdetr-medium-q8_0.gguf
93
+ 7a6c0da4b1b1070986f30582eb8b1b7439b386b15a2171707346eb7c402cca60 rfdetr-medium-q4_K.gguf
94
+ ```
95
+
96
+ ## License
97
+
98
+ Apache-2.0 — matches the upstream [rfdetr](https://github.com/roboflow/rf-detr) license.