Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -26,6 +26,12 @@ All variants take and return fp32 tensors — swap the `.pte` file, keep your ap
|
|
| 26 |
\*Mac arm64, single process, median of 10 — a reference point for relative cost
|
| 27 |
only, not a device number (torch eager fp32 on the same machine: 392.2 ms).
|
| 28 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
### Precisions that did not earn a slot
|
| 30 |
|
| 31 |
- **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.
|
|
|
|
| 26 |
\*Mac arm64, single process, median of 10 — a reference point for relative cost
|
| 27 |
only, not a device number (torch eager fp32 on the same machine: 392.2 ms).
|
| 28 |
|
| 29 |
+
### Checked in the task's own units
|
| 30 |
+
|
| 31 |
+
Correlation is a first filter. These are the numbers that decide:
|
| 32 |
+
|
| 33 |
+
- **int8** — mask IoU 0.963 median against fp32 (worst 0.927 of five images). The disagreement is boundary pixels, which is what int8 costs on a cutout model.
|
| 34 |
+
|
| 35 |
### Precisions that did not earn a slot
|
| 36 |
|
| 37 |
- **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.
|