GitLab

PiperOrigin-RevId: 398596089

using 5d tensors to avoid issue when converted to TfLite model.

PiperOrigin-RevId: 398301856

PiperOrigin-RevId: 397358689

PiperOrigin-RevId: 395297294

PiperOrigin-RevId: 394568069

PiperOrigin-RevId: 394512536

PiperOrigin-RevId: 394505975

CenterNet meta arch to avoid using scatter_nd and tf.cond.

Also, use static_or_dynamic_map_fn in the convert_strided_predictions_to_normalized_keypoints function to avoid introducing cycles in the graph that causes the TPU converter errors.

PiperOrigin-RevId: 394152521

runs much slower in the browser.

PiperOrigin-RevId: 394148346

PiperOrigin-RevId: 393480597

The order of the input tensors is incorrect and therefore the sample notebook does not run.

The sample code has the order of the input tensors as `boxes, classes, scores, num_detections`, but actually the order should be `scores, boxes,  num_detections, classes`

After correcting the indexes of the input tensor, the sample notebook runs.

PiperOrigin-RevId: 392061572

  * Replaced the deprecated `to_int/to_float` in `center_net_meta_arch.py` with `tf.cast()` (following the warning).
  * Added `max_norm_coord` argument to CenterNet source code to allow controlling the maximum coordinate value to be considered, to allow for the wrap-around special case for RxR data.

PiperOrigin-RevId: 391319766

PiperOrigin-RevId: 391079006

true_image_shape is not required and can be inferred by the prediction tensors.

PiperOrigin-RevId: 390421432

PiperOrigin-RevId: 388508391

PiperOrigin-RevId: 385646058

PiperOrigin-RevId: 385342213

PiperOrigin-RevId: 385224486

Made changes in `center_net_meta_arch.py` for the particular use case in Marky Landmark Detection, i.e.
1. Accessing an access-protected field in the model;
2. Traverse layers to set trainability.

PiperOrigin-RevId: 385136419

PiperOrigin-RevId: 385042521

PiperOrigin-RevId: 383699499

PiperOrigin-RevId: 383651497

PiperOrigin-RevId: 383469336

PiperOrigin-RevId: 382773317

PiperOrigin-RevId: 382236941

PiperOrigin-RevId: 381366438

This allows the exported TF Lite MoveNet model to be able to run on device GPU
using OpenCL/OpenGL.

PiperOrigin-RevId: 381343791

over new parameters in the CenterNet model.

PiperOrigin-RevId: 381331623

PiperOrigin-RevId: 381307288

PiperOrigin-RevId: 379898918

PiperOrigin-RevId: 379828181

PiperOrigin-RevId: 379817535

1) Added the score-to-scaled-distance ranking method.
2) Added the gaussian-weighted ranking method.
3) Added the hyperparameter used in the rescoring function.

PiperOrigin-RevId: 379646603

1. Make api_template always use third_party/py/keras (or keras PIP package in OSS).
2. API tests are updated to ignore tensorflow/python/keras end points.
3. All Keras-related code will be removed in a follow up change. Keras-specific APIs are tracked in google3/third_party/py/keras/api/golden/
4. There are a few APIs (optimizer/initializer) that are exposed as both tf and keras APIs. Their pbtxt are updated to reflect the latest class path change.

Once this change is submitted, the current code in tensorflow/python/keras will become a stale copy, and will be removed soon. No more changes should be made to tensorflow/python/keras.

PiperOrigin-RevId: 379353790

PiperOrigin-RevId: 378977617

Add notebook and update ODT TFLite conversion doc to show how to make the model compatible with TFLite Task Library.

PiperOrigin-RevId: 378313314

region for the keypoint heatmap if the instance's keypoint visibility is 0.

PiperOrigin-RevId: 378183175

PiperOrigin-RevId: 378154667