Flor commited on
Commit
bbffb60
·
verified ·
1 Parent(s): 12e32fe

card: update from fastforks 24802779

Browse files
Files changed (1) hide show
  1. README.md +14 -14
README.md CHANGED
@@ -18,7 +18,7 @@ Selective-IGEV (Wang et al. 2024, Selective-Stereo): IGEV-Stereo with selective
18
 
19
  This repo is the model card only: it holds no weights. fastforks downloads the authors' weights from the links below and checks them against the SHA-256 in the model card. Code: <https://github.com/kleinicke/fastforks>. Kernels for the `kernels` backend: <https://huggingface.co/fastforks/kernels>. Original code: <https://github.com/Windsrain/Selective-Stereo> at `c39acbfef9fb`.
20
 
21
- Task: stereo. Tier: experimental. Outputs: `disparity`. Backends: `portable`, `kernels`.
22
 
23
  Input: a rectified left / right image pair. Output: a disparity map (pixels, left view).
24
 
@@ -67,9 +67,15 @@ Tested up to (portable, fp16, RTX 5060 Ti 16 GB): `selective-igev-eth3d-finetune
67
 
68
  ### torch.compile and TensorRT
69
 
70
- | model | input | portable ms | compile ms | compile s (cold) | TensorRT fp16 ms | ONNX + build s | kernels ms | peak MiB (alloc / reserved): portable | compile | TensorRT (device) |
71
- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
72
- | `selective-igev-eth3d-finetune` | 994x1482 | 998 | 690.7 | 125.8 | 460.8 | 256 | 349 | 1238 / 1488 | 919 / 1088 | 1148 |
 
 
 
 
 
 
73
 
74
  ### Fixed-size speed and peak memory, all models
75
 
@@ -90,23 +96,17 @@ Tested up to (portable, fp16, RTX 5060 Ti 16 GB): `selective-igev-eth3d-finetune
90
  | `selective-igev-sceneflow` | portable | 190.6 ms / 282 / 362 | 575.4 ms / 769 / 1036 | 1420.3 ms / 1690 / 2030 | 8917.6 ms (n=3) / 10549 / 14320 |
91
  | `selective-igev-sceneflow` | kernels | 82.5 ms / 260 / 368 | 213.5 ms / 693 / 1010 | 473.3 ms / 1502 / 2046 | 2808.4 ms (n=3) / 6928 / 8800 |
92
 
93
- ### Apple silicon (MPS)
94
-
95
- | model | `portable` | `kernels` | speed-up |
96
- | --- | --- | --- | --- |
97
- | `selective-igev-sceneflow` | 2292 ms / 1727 MiB (float16) | 1878 ms / 1704 MiB | 1.22x |
98
-
99
  Fixed-size cells: time (ms) / peak allocated / peak reserved memory (MiB).
100
 
101
  - `selective-igev-eth3d-train`, `selective-igev-kitti12`, `selective-igev-kitti15`, `selective-igev-middlebury-finetune`, `selective-igev-middlebury-train`, `selective-igev-sceneflow`: same network and settings as `selective-igev-eth3d-finetune`, measured once (rows not repeated).
102
 
103
- Hardware, protocol and caveats: [benchmarks](https://github.com/kleinicke/fastforks/blob/main/docs/benchmarks.md) (RTX 5060 Ti 16 GB, sm_120). Original = the authors' code at the precision of portable (fp16, or `(fp32)` where fp16 is unsupported or not finite); original fp32 = its default fp32 run. Peak memory is the torch allocator's.
104
 
105
  ## Accuracy
106
 
107
  ### On public datasets
108
 
109
- Original = the authors' code, same checkpoint and protocol.
110
 
111
  | variant | dataset | original fp32 | original fp16 | portable | kernels | note |
112
  |---|---|---|---|---|---|---|
@@ -125,7 +125,7 @@ Original = the authors' code, same checkpoint and protocol.
125
  | `selective-igev-eth3d-train` | ETH3D two-view train, bad > 1 px % / EPE | 0.78 / 0.137 | 0.78 / 0.137 | 0.78 / 0.137 | 0.78 / 0.137 | trained on this dataset's public images; not a zero-shot result |
126
  | `selective-igev-eth3d-train` | Middlebury 2014 trainingH, bad > 2 px % / EPE | 7.96 / 2.158 | 7.97 / 2.153 | 7.99 / 2.161 | 8.01 / 2.158 | |
127
 
128
- Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-05.
129
 
130
  ### Parity with the original code
131
 
@@ -152,4 +152,4 @@ Install with `pip install fastforks` (see <https://github.com/kleinicke/fastfork
152
 
153
  `stereo.load("selective-igev-eth3d-finetune", mode="kernels")` runs our CUDA kernels (`igev`, `cutlass_conv`, `cudnn_hacc`). It needs a prebuilt kernel for your GPU architecture (downloaded from <https://huggingface.co/fastforks/kernels> on first use) and a recent NVIDIA driver; if there is none it raises `BackendUnavailableError` with the reason and never falls back silently. `mode="auto"` picks `kernels` when it can run and logs the choice.
154
 
155
- <!-- staged from fastforks feccac41 (2026-11-06) -->
 
18
 
19
  This repo is the model card only: it holds no weights. fastforks downloads the authors' weights from the links below and checks them against the SHA-256 in the model card. Code: <https://github.com/kleinicke/fastforks>. Kernels for the `kernels` backend: <https://huggingface.co/fastforks/kernels>. Original code: <https://github.com/Windsrain/Selective-Stereo> at `c39acbfef9fb`.
20
 
21
+ Task: stereo. Tier: stable. Outputs: `disparity`. Backends: `portable`, `kernels`.
22
 
23
  Input: a rectified left / right image pair. Output: a disparity map (pixels, left view).
24
 
 
67
 
68
  ### torch.compile and TensorRT
69
 
70
+ | model | input | time: portable | compile | TensorRT fp16 | kernels | preparation: compile (cold first call) | ONNX + engine build | peak MiB (alloc / reserved): portable | compile | kernels | TensorRT (engine + torch) |
71
+ | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
72
+ | `selective-igev-eth3d-finetune` | 994x1482 | 998 ms | 691 ms | 461 ms | 349 ms | 126 s | 256 s | 1238 / 1488 | 919 / 1088 | 1109 / 1390 | 1279 / 1316 |
73
+ | `selective-igev-eth3d-train` | 994x1482 | 998 ms | 691 ms | 462 ms | 349 ms | 127 s | 260 s | 1238 / 1488 | 915 / 1028 | 1109 / 1390 | 1278 / 1334 |
74
+ | `selective-igev-kitti12` | 994x1482 | 998 ms | 691 ms | 462 ms | 349 ms | 125 s | 255 s | 1238 / 1488 | 915 / 1028 | 1109 / 1390 | 1278 / 1334 |
75
+ | `selective-igev-kitti15` | 994x1482 | 998 ms | 691 ms | 462 ms | 349 ms | 127 s | 256 s | 1238 / 1488 | 915 / 1028 | 1109 / 1390 | 1278 / 1334 |
76
+ | `selective-igev-middlebury-finetune` | 994x1482 | 998 ms | 692 ms | 462 ms | 349 ms | 127 s | 258 s | 1238 / 1488 | 915 / 1028 | 1109 / 1390 | 1280 / 1316 |
77
+ | `selective-igev-middlebury-train` | 994x1482 | 998 ms | 692 ms | 462 ms | 349 ms | 128 s | 256 s | 1238 / 1488 | 915 / 1028 | 1109 / 1390 | 1278 / 1334 |
78
+ | `selective-igev-sceneflow` | 994x1482 | 998 ms | 692 ms | 463 ms | 349 ms | 127 s | 255 s | 1238 / 1488 | 915 / 1028 | 1109 / 1390 | 1278 / 1334 |
79
 
80
  ### Fixed-size speed and peak memory, all models
81
 
 
96
  | `selective-igev-sceneflow` | portable | 190.6 ms / 282 / 362 | 575.4 ms / 769 / 1036 | 1420.3 ms / 1690 / 2030 | 8917.6 ms (n=3) / 10549 / 14320 |
97
  | `selective-igev-sceneflow` | kernels | 82.5 ms / 260 / 368 | 213.5 ms / 693 / 1010 | 473.3 ms / 1502 / 2046 | 2808.4 ms (n=3) / 6928 / 8800 |
98
 
 
 
 
 
 
 
99
  Fixed-size cells: time (ms) / peak allocated / peak reserved memory (MiB).
100
 
101
  - `selective-igev-eth3d-train`, `selective-igev-kitti12`, `selective-igev-kitti15`, `selective-igev-middlebury-finetune`, `selective-igev-middlebury-train`, `selective-igev-sceneflow`: same network and settings as `selective-igev-eth3d-finetune`, measured once (rows not repeated).
102
 
103
+ Hardware, protocol and caveats: [benchmarks](https://github.com/kleinicke/fastforks/blob/main/docs/benchmarks.md) (RTX 5060 Ti 16 GB, sm_120). Original = the original code named above at the precision of portable (fp16, or `(fp32)` where fp16 is unsupported or not finite); original fp32 = its default fp32 run. Peak memory is the torch allocator's.
104
 
105
  ## Accuracy
106
 
107
  ### On public datasets
108
 
109
+ Original = the original code named under Details, same checkpoint and protocol.
110
 
111
  | variant | dataset | original fp32 | original fp16 | portable | kernels | note |
112
  |---|---|---|---|---|---|---|
 
125
  | `selective-igev-eth3d-train` | ETH3D two-view train, bad > 1 px % / EPE | 0.78 / 0.137 | 0.78 / 0.137 | 0.78 / 0.137 | 0.78 / 0.137 | trained on this dataset's public images; not a zero-shot result |
126
  | `selective-igev-eth3d-train` | Middlebury 2014 trainingH, bad > 2 px % / EPE | 7.96 / 2.158 | 7.97 / 2.153 | 7.99 / 2.161 | 8.01 / 2.158 | |
127
 
128
+ Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 468fba10, 2026-11-05.
129
 
130
  ### Parity with the original code
131
 
 
152
 
153
  `stereo.load("selective-igev-eth3d-finetune", mode="kernels")` runs our CUDA kernels (`igev`, `cutlass_conv`, `cudnn_hacc`). It needs a prebuilt kernel for your GPU architecture (downloaded from <https://huggingface.co/fastforks/kernels> on first use) and a recent NVIDIA driver; if there is none it raises `BackendUnavailableError` with the reason and never falls back silently. `mode="auto"` picks `kernels` when it can run and logs the choice.
154
 
155
+ <!-- staged from fastforks 24802779 (2026-11-15) -->