GitLab

PiperOrigin-RevId: 370255238

logic doesn't depend on the per_keypoint_offset value and use
get_batch_predictions_from_indices to be consistent with other prediction
heads.

PiperOrigin-RevId: 369925004

PiperOrigin-RevId: 369749466

* OD tutorial colab working

* fixing install instructions link

Surgical update to CenterNet keypoint postprocessing to use bounding boxes if present. This is shown to reduce keypoint snapping across individuals.

PiperOrigin-RevId: 369432018

Call set_keypoint_visibilities on the non-expanded versions of the detection and groundtruth keypoints. set_keypoint_visibilities expects a rank-3 tensor, and was being provided a rank-4 tensor. This had the unintended effect of creating a keypoint visibilities tensor of the wrong shape, resulting in only 2 keypoints being visualized.

PiperOrigin-RevId: 368458273

Fix the typo-ed name "remove_unecessary_ema" -> "remove_unnecessary_ema" and also re-implement it with deep dictionary copy.

PiperOrigin-RevId: 368338501

PiperOrigin-RevId: 368217758

PiperOrigin-RevId: 368215027

This replaces symbols in tf.keras.mixed_precision.experimental with the corresponding nonexperimental symbols. In some cases, passing a Policy is replaced with passing a policy name for conciseness.

Additionally, for the Shakespeare model, the loss_scale flag is removed, since supporting it with the nonexperimental API is slightly more verbose and it is recommended users use the default loss scale.

PiperOrigin-RevId: 368123944

PiperOrigin-RevId: 368024164

Rewrite the GIOU loss op in matrix form to avoid the map_fn. map_fn could be prohibitively costly when dealing with large number of bounding boxes.

PiperOrigin-RevId: 366854232

PiperOrigin-RevId: 366850188

PiperOrigin-RevId: 366395416

PiperOrigin-RevId: 366388829

PiperOrigin-RevId: 366385685

PiperOrigin-RevId: 366382220

PiperOrigin-RevId: 366134194

PiperOrigin-RevId: 365913722

PiperOrigin-RevId: 365899795

This CL register a pre-defined GIOU loss to the losses.proto and losses_builder so that the GIOU loss can be correctly called & built when GIOU loss is specified in the loss config file.

PiperOrigin-RevId: 365891022

PiperOrigin-RevId: 365683210

PiperOrigin-RevId: 365177695

PiperOrigin-RevId: 364924351

PiperOrigin-RevId: 364895608

PiperOrigin-RevId: 364731599

PiperOrigin-RevId: 364671967

Updated the single instance postprocessing logics such that it uses the score to distance ratio scoring method. The new logics also avoid using expensive ops, e.g. reduce_max/maximum.

PiperOrigin-RevId: 364607976

PiperOrigin-RevId: 364388006

PiperOrigin-RevId: 364363163

PiperOrigin-RevId: 364349645

PiperOrigin-RevId: 363975731

PiperOrigin-RevId: 363794263

PiperOrigin-RevId: 363700908

use case which uses the keypoint location to determine the object center
location and heatmap.

PiperOrigin-RevId: 363522414

postprocessing logic.

PiperOrigin-RevId: 362814113

PiperOrigin-RevId: 362609612

Updating how keypoints are rescored in CenterNet. Previously, all keypoint scores were averaged. Now, any keypoints with confidence score 0 (e.g. out of frame keypoints) are not averaged.

PiperOrigin-RevId: 362383830

PiperOrigin-RevId: 362353183