The vvas_xinfer element selects the tensor format from three values:
model-format from the JSON configuration, the model tensor layout, and the
tensor data type. GStreamer logs and caps use the names in the last column.
model-format |
Tensor layout |
Data type |
GStreamer format |
|---|---|---|---|
RGB |
NCHW |
FLOAT32 |
|
RGB |
NHWC |
FLOAT32 |
|
RGB |
NCHW |
INT8 |
|
RGB |
NHWC |
INT8 |
|
RGB |
HCWNC4 |
INT8 |
|
RGB |
NCHW |
BF16 |
|
RGB |
NHWC |
BF16 |
|
RGB |
HCWNC4 |
BF16 |
|
RGB |
NCHW |
FP16 |
|
RGB |
NHWC |
FP16 |
|
RGB |
HCWNC4 |
FP16 |
|
RGB |
HCWNC8 |
INT8 |
|
RGB |
HCWNC8 |
BF16 |
|
RGB |
HCWNC8 |
FP16 |
|
BGR |
NHWC |
FLOAT32 |
|
BGR |
NHWC |
INT8 |
|
BGR |
NCHW |
BF16 |
|
BGR |
NHWC |
BF16 |
|
BGR |
HCWNC4 |
BF16 |
|
BGR |
NCHW |
FP16 |
|
BGR |
NHWC |
FP16 |
|
BGR |
HCWNC4 |
FP16 |
|
vvas_xinfer does not select BGR with NCHW for FLOAT32 or INT8 models.
GRAY_BF16, GRAY_FP16, and GRAY_FLOAT may appear in caps for tensor
compatibility, but they are not selected for RGB or BGR model configurations.