In Hailo DFC, is it supported to use non-image inputs, such as feature tensors or token embeddings, as network inputs?
Specifically, I have ONNX models whose inputs are rank-3 tensors with shape (Batch, Tokens, Channels) (e.g. (4, 256, 1536)), representing precomputed features rather than images.
I would like to:
- Prevent the compiler from treating these inputs as image/RGB inputs
- Disable any automatic RGB or image-related preprocessing (e.g. TF RGB → Hailo RGB conversion)
- Explicitly mark these inputs as raw feature tensors
What is the recommended way to define such inputs in DFC?
-
Should net_input_format with [Dims.BATCH, Dims.WIDTH, Dims.CHANNELS] be sufficient?
-
Is there an explicit flag or ALLS directive to force an input to be treated as a non-image (raw) input?
