mlboydaisuke commited on
Commit
028f720
·
verified ·
1 Parent(s): 9a61a13

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +12 -4
README.md CHANGED
@@ -9,7 +9,7 @@ tags:
9
  base_model:
10
  - depth-anything/Depth-Anything-V2-Small-hf
11
  ---
12
- # depth_anything_v2_small — ExecuTorch XNNPACK
13
 
14
  - **Source**: depth-anything/Depth-Anything-V2-Small-hf
15
  - **License**: Apache-2.0
@@ -20,15 +20,23 @@ base_model:
20
 
21
  All variants take and return fp32 tensors — swap the `.pte` file, keep your app code.
22
 
23
- | precision | file | size (MB) | parity vs fp32 eager (worst corr) | Mac median (ms)* |
24
  |-----------|------|-----------|------------------------------------|------------------|
25
  | fp32 | `depth_anything_v2_small_xnnpack_fp32.pte` | 99.0 | 1.000000 | 167.3 |
26
  | fp16 | `depth_anything_v2_small_xnnpack_fp16.pte` | 55.5 | 0.999992 | 289.7 |
 
 
 
 
 
 
 
 
27
 
28
  \*Mac arm64, single process, median of 10 — a reference point for relative cost
29
  only, not a device number (torch eager fp32 on the same machine: 85.0 ms).
30
 
31
- ### Precisions that did not earn a slot
32
 
33
  - **int8 (dynamic) is not shipped**: measured in the units that matter for this model — fraction of pixels within 1.25x of the fp32 depth: median 0.9941 over 10 real images, worst 0.9748.
34
 
@@ -45,5 +53,5 @@ XNNPACK delegate coverage (fp32): 73.4% (482/657 ops); ops left on the portable
45
 
46
  ## Conversion
47
 
48
- torch.export -> to_edge_transform_and_lower(XnnpackPartitioner) -> .pte
49
  (conversion scripts: [executorch-models](https://github.com/john-rocky/executorch-models))
 
9
  base_model:
10
  - depth-anything/Depth-Anything-V2-Small-hf
11
  ---
12
+ # depth_anything_v2_small — ExecuTorch
13
 
14
  - **Source**: depth-anything/Depth-Anything-V2-Small-hf
15
  - **License**: Apache-2.0
 
20
 
21
  All variants take and return fp32 tensors — swap the `.pte` file, keep your app code.
22
 
23
+ | build | file | size (MB) | parity vs fp32 eager (worst corr) | Mac median (ms)* |
24
  |-----------|------|-----------|------------------------------------|------------------|
25
  | fp32 | `depth_anything_v2_small_xnnpack_fp32.pte` | 99.0 | 1.000000 | 167.3 |
26
  | fp16 | `depth_anything_v2_small_xnnpack_fp16.pte` | 55.5 | 0.999992 | 289.7 |
27
+ | Core ML (fp16, iOS) | `depth_anything_v2_small_coreml_all.pte` | 50.2 | 0.999992 | 48.1 |
28
+
29
+
30
+ The Core ML build is the same graph lowered to Apple's Neural Engine instead of
31
+ XNNPACK, which is CPU-only. On an iPhone 17 Pro, Depth-Anything-V2-Small runs
32
+ 500.8 ms through XNNPACK and 42.7 ms through Core ML, at half the file size. It
33
+ computes in fp16 and is iOS-only; the XNNPACK files stay the portable option and
34
+ are what runs on Android.
35
 
36
  \*Mac arm64, single process, median of 10 — a reference point for relative cost
37
  only, not a device number (torch eager fp32 on the same machine: 85.0 ms).
38
 
39
+ ### Builds that did not earn a slot
40
 
41
  - **int8 (dynamic) is not shipped**: measured in the units that matter for this model — fraction of pixels within 1.25x of the fp32 depth: median 0.9941 over 10 real images, worst 0.9748.
42
 
 
53
 
54
  ## Conversion
55
 
56
+ torch.export -> to_edge_transform_and_lower(partitioner) -> .pte
57
  (conversion scripts: [executorch-models](https://github.com/john-rocky/executorch-models))