Support arbitrary data bits
dbericat opened this issue · 0 comments
dbericat commented
Is your feature request related to a problem? Please describe.
- Buffers are not floating point buffers (integers or less than integers 8-4-4 or even 4-2-2)
- Limitations in how data flows around in GStreamer (video format, can’t pass floating points numbers, no point in transforming before inference)
- Transforms -> inference -> transforms flow can’t be done. Hast to be done as G streamer data types.
Describe the solution you'd like
- Goal: Read into CUDA tensors
- DS fix: parameters on inference that handles offset
Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.
Additional context
- Need use cases: current models from KCL had 1 channel input. Had to retrain to have 3 channels
- Select from 2-3 # of channels down to 1
- How do we insert pre-transforms
- Can’t do it in a separate component.
- Enhance DS inference server