GitLab

PiperOrigin-RevId: 372426423

interpolation method is configurable.

PiperOrigin-RevId: 372198977

PiperOrigin-RevId: 371929331

for the object center head parameters in CenterNet.

PiperOrigin-RevId: 371783912

PiperOrigin-RevId: 371037598

Fix explicit padding when pool residuals are used. Match the pool residual convolution to the depthwise convolution to ensure the output tensors of both are of the same size.

PiperOrigin-RevId: 370971019

PiperOrigin-RevId: 370903487

PiperOrigin-RevId: 370720011

PiperOrigin-RevId: 370575507

CenterNetMetaArch. It is to make sure the model can be post-quantizable and
doesn't generate calibration errors during quantization.

PiperOrigin-RevId: 370528566

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