Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -7,7 +7,7 @@ tags:
|
|
| 7 |
- on-device
|
| 8 |
- image-segmentation
|
| 9 |
---
|
| 10 |
-
# dis_isnet — ExecuTorch
|
| 11 |
|
| 12 |
- **Source**: xuebinqin/DIS + NimaBoscarino/IS-Net_DIS-general-use weights
|
| 13 |
- **License**: Apache-2.0
|
|
@@ -18,14 +18,22 @@ tags:
|
|
| 18 |
|
| 19 |
All variants take and return fp32 tensors — swap the `.pte` file, keep your app code.
|
| 20 |
|
| 21 |
-
|
|
| 22 |
|-----------|------|-----------|------------------------------------|------------------|
|
| 23 |
| fp32 | `dis_isnet_xnnpack_fp32.pte` | 176.1 | 1.000000 | 123.4 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
|
| 25 |
\*Mac arm64, single process, median of 10 — a reference point for relative cost
|
| 26 |
only, not a device number (torch eager fp32 on the same machine: 364.6 ms).
|
| 27 |
|
| 28 |
-
###
|
| 29 |
|
| 30 |
- **fp16 is not shipped**: worst-output corr 0.986 against fp32 eager, below the 0.995 bar for this precision. The file converts and runs; the numbers do not hold up, so it is left out rather than shipped with a warning.
|
| 31 |
- **int8 is not shipped**: measured in the units that matter for this model — mask IoU at 0.5: median 0.9106 over 10 real images, worst 0.4647.
|
|
@@ -43,5 +51,5 @@ XNNPACK delegate coverage (fp32): 100.0% (467/467 ops)
|
|
| 43 |
|
| 44 |
## Conversion
|
| 45 |
|
| 46 |
-
torch.export -> to_edge_transform_and_lower(
|
| 47 |
(conversion scripts: [executorch-models](https://github.com/john-rocky/executorch-models))
|
|
|
|
| 7 |
- on-device
|
| 8 |
- image-segmentation
|
| 9 |
---
|
| 10 |
+
# dis_isnet — ExecuTorch
|
| 11 |
|
| 12 |
- **Source**: xuebinqin/DIS + NimaBoscarino/IS-Net_DIS-general-use weights
|
| 13 |
- **License**: Apache-2.0
|
|
|
|
| 18 |
|
| 19 |
All variants take and return fp32 tensors — swap the `.pte` file, keep your app code.
|
| 20 |
|
| 21 |
+
| build | file | size (MB) | parity vs fp32 eager (worst corr) | Mac median (ms)* |
|
| 22 |
|-----------|------|-----------|------------------------------------|------------------|
|
| 23 |
| fp32 | `dis_isnet_xnnpack_fp32.pte` | 176.1 | 1.000000 | 123.4 |
|
| 24 |
+
| Core ML (fp16, iOS) | `dis_isnet_coreml_all.pte` | 89.0 | 0.999984 | 29.2 |
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
The Core ML build is the same graph lowered to Apple's Neural Engine instead of
|
| 28 |
+
XNNPACK, which is CPU-only. On an iPhone 17 Pro, Depth-Anything-V2-Small runs
|
| 29 |
+
500.8 ms through XNNPACK and 42.7 ms through Core ML, at half the file size. It
|
| 30 |
+
computes in fp16 and is iOS-only; the XNNPACK files stay the portable option and
|
| 31 |
+
are what runs on Android.
|
| 32 |
|
| 33 |
\*Mac arm64, single process, median of 10 — a reference point for relative cost
|
| 34 |
only, not a device number (torch eager fp32 on the same machine: 364.6 ms).
|
| 35 |
|
| 36 |
+
### Builds that did not earn a slot
|
| 37 |
|
| 38 |
- **fp16 is not shipped**: worst-output corr 0.986 against fp32 eager, below the 0.995 bar for this precision. The file converts and runs; the numbers do not hold up, so it is left out rather than shipped with a warning.
|
| 39 |
- **int8 is not shipped**: measured in the units that matter for this model — mask IoU at 0.5: median 0.9106 over 10 real images, worst 0.4647.
|
|
|
|
| 51 |
|
| 52 |
## Conversion
|
| 53 |
|
| 54 |
+
torch.export -> to_edge_transform_and_lower(partitioner) -> .pte
|
| 55 |
(conversion scripts: [executorch-models](https://github.com/john-rocky/executorch-models))
|