Closed IMvision12 closed 1 year ago
OD does not require ragged Tensors -- in many cases we pad to a dense representation.
JAX does not include support for ragged arrays, and I'm not aware of a strong Torch convention for this either, so for now ragged support is limited to the backend that offers them natively.
Object detection models in KerasCV necessitate the use of ragged tensors. While TensorFlow as a backend supports this but jax and pytorch backend don't support it.
@ianstenbit @fchollet @tirthasheshpatel