Port CenterPillar to Keras Core #2003
Conversation
|
/gcbrun |
|
/gcbrun |
There was a problem hiding this comment.
@ianstenbit I thought this model wasn't possible to port due to custom tf ops and I obviously see a lot of TF code remaining. Is the idea that this is a TensorFlow-only model that works with Keras Core? If so, is this something we can flag for users?
There was a problem hiding this comment.
This file still has a LOT of tensorflow code, is that intentional?
There was a problem hiding this comment.
Yep! This lives outside of the model and runs in tf.data
| def point_to_voxel_coord( | ||
| point_xyz: tf.Tensor, voxel_size: Sequence[float], dtype=tf.int32 | ||
| ) -> tf.Tensor: | ||
| def point_to_voxel_coord(point_xyz, voxel_size, dtype=tf.int32): |
There was a problem hiding this comment.
This function also has a lot of TF
There was a problem hiding this comment.
(Same here -- this is for tf.data)
This does work with other backends, but requires data encoding in a tf.data pipeline, so that code is all still written in TF. |
|
/gcbrun |
|
/gcbrun |
|
/gcbrun |
* Port CenterPillar to Keras Core * Switch back to backbone API * Restructure center_pillar.py * Undo some mistaken changes * More fixes -- tests currently failing * Test passing across backends * Remove local segment_max op * Finish port + tests passing * Fix tests + format * Remove numpy behavior * Moar test fixes * Use smaller backbone * Review comments * Add 3D OD models to multi-backend CI
TODOs (all done now -- sending for review)
_coreversions of voxel_utilssegment_maxkeras_core.backend.name_scopecenter_pillar.py