Group backend precision notes under a shared heading
Browse files
README.md
CHANGED
|
@@ -35,7 +35,7 @@ opaque RGBA grayscale mask; use its foreground intensity as an alpha mask when
|
|
| 35 |
compositing the original image. Fine structures can be missed. FP16 favors
|
| 36 |
speed and may shift mask boundaries; FP32 preserves the reference more closely.
|
| 37 |
|
| 38 |
-
##
|
| 39 |
|
| 40 |
The INT8 variant selectively quantizes four expensive convolutions with static
|
| 41 |
INT8 activations and per-channel INT8 weights. Sensitive layers and skip
|
|
|
|
| 35 |
compositing the original image. Fine structures can be missed. FP16 favors
|
| 36 |
speed and may shift mask boundaries; FP32 preserves the reference more closely.
|
| 37 |
|
| 38 |
+
## Precision notes
|
| 39 |
|
| 40 |
The INT8 variant selectively quantizes four expensive convolutions with static
|
| 41 |
INT8 activations and per-channel INT8 weights. Sensitive layers and skip
|