GitLab

250447559  by Zhichao Lu:

    Update expected files format for Instance Segmentation challenge:
    - add fields ImageWidth, ImageHeight and store the values per prediction
    - as mask, store only encoded image and assume its size is ImageWidth x ImageHeight

--
250402780  by rathodv:

    Fix failing Mask R-CNN TPU convergence test.

    Cast second stage prediction tensors from bfloat16 to float32 to prevent errors in third target assignment (Mask Prediction) - Concat with different types bfloat16 and bfloat32 isn't allowed.

--
250300240  by Zhichao Lu:

    Addion Open Images Challenge 2019 object detection and instance segmentation
    support into Estimator framework.

--
249944839  by rathodv:

    Modify exporter.py to add multiclass score nodes in exported inference graphs.

--
249935201  by rathodv:

    Modify postprocess methods to preserve multiclass scores after non max suppression.

--
249878079  by Zhichao Lu:

    This CL slightly refactors some Object Detection helper functions for data creation, evaluation, and groundtruth providing.

    This will allow the eager+function custom loops to share code with the existing estimator training loops.

    Concretely we make the following changes:
    1. In input creation we separate dataset-creation into top-level helpers, and allow it to optionally accept a pre-constructed model directly instead of always creating a model from the config just for feature preprocessing.

    2. In coco evaluation we split the update_op creation into its own function, which the custom loops will call directly.

    3. In model_lib we move groundtruth providing/ datastructure munging into a helper function

    4. For now we put an escape hatch in `_summarize_target_assignment` when executing in tf v2.0 behavior because the summary apis used only work w/ tf 1.x

--
249673507  by rathodv:

    Use explicit casts instead of tf.to_float and tf.to_int32 to avoid warnings.

--
249656006  by Zhichao Lu:

    Add named "raw_keypoint_locations" node that corresponds with the "raw_box_locations" node.

--
249651674  by rathodv:

    Keep proposal boxes in float format. MatMulCropAndResize can handle the type even when feature themselves are bfloat16s.

--
249568633  by rathodv:

    Support q > 1 in class agnostic NMS.
    Break post_processing_test.py into 3 separate files to avoid linter errors.

--
249535530  by rathodv:

    Update some deprecated arguments to tf ops.

--
249368223  by rathodv:

    Modify MatMulCropAndResize to use MultiLevelRoIAlign method and move the tests to spatial_transform_ops.py module.

    This cl establishes that CropAndResize and RoIAlign are equivalent and only differ in the sampling point grid within the boxes. CropAndResize uses a uniform size x size point grid such that the corner points exactly overlap box corners, while RoiAlign divides boxes into size x size cells and uses their centers as sampling points. In this cl, we switch MatMulCropAndResize to use the MultiLevelRoIAlign implementation with `align_corner` option as MultiLevelRoIAlign implementation is more memory efficient on TPU when compared to the original MatMulCropAndResize.

--
249337338  by chowdhery:

    Add class-agnostic non-max-suppression in post_processing

--
249139196  by Zhichao Lu:

    Fix positional argument bug in export_tflite_ssd_graph

--
249120219  by Zhichao Lu:

    Add evaluator for computing precision limited to a given recall range.

--
249030593  by Zhichao Lu:

    Evaluation util to run segmentation and detection challenge evaluation.

--
248554358  by Zhichao Lu:

    This change contains the auxiliary changes required for TF 2.0 style training with eager+functions+dist strat loops, but not the loops themselves.

    It includes:
    - Updates to shape usage to support both tensorshape v1 and tensorshape v2
    - A fix to FreezableBatchNorm to not override the `training` arg in call when `None` was passed to the constructor (Not an issue in the estimator loops but it was in the custom loops)
    - Puts some constants in init_scope so they work in eager + functions
    - Makes learning rate schedules return a callable in eager mode (required so they update when the global_step changes)
    - Makes DetectionModel a tf.module so it tracks variables (e.g. ones nested in layers)
    - Removes some references to `op.name` for some losses and replaces it w/ explicit names
    - A small part of the change to allow the coco evaluation metrics to work in eager mode

--
248271226  by rathodv:

    Add MultiLevel RoIAlign op.

--
248229103  by rathodv:

    Add functions to 1. pad features maps 2. ravel 5-D indices

--
248206769  by rathodv:

    Add utilities needed to introduce RoI Align op.

--
248177733  by pengchong:

    Internal changes

--
247742582  by Zhichao Lu:

    Open Images Challenge 2019 instance segmentation metric: part 2

--
247525401  by Zhichao Lu:

    Update comments on max_class_per_detection.

--
247520753  by rathodv:

    Add multilevel crop and resize operation that builds on top of matmul_crop_and_resize.

--
247391600  by Zhichao Lu:

    Open Images Challenge 2019 instance segmentation metric

--
247325813  by chowdhery:

    Quantized MobileNet v2 SSD FPNLite config with depth multiplier 0.75

--

PiperOrigin-RevId: 250447559

250779087  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Reduce BERT Perfzero benchmark test training steps.

--

PiperOrigin-RevId: 250779087

* Support pure eager execution in ResNet50

* Use smaller batch size

250713045  by hongkuny<hongkuny@google.com>:

    TPU util

--

PiperOrigin-RevId: 250713045

250606180  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Fix BERT benchamrk test errors.

--
250589623  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Change BERT benchmark test pretrained checkpoint url.

--
250587892  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Fix error in BERT custom training loop checkpoint restoration.

--
250577163  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Add logic to inject callback that measures performance in BERT custom training
    loop.

--
250529526  by hongkuny<hongkuny@google.com>:

    Internal clean up

--
250428976  by hongkuny<hongkuny@google.com>:

    Internal change

250415383  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Add min/max value to BERT classifier benchmark test.

--
250376246  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Add benchmark performance test to run BERT on multiple numbers of GPUs.

--

PiperOrigin-RevId: 250606180

* Put all python dependencies into one line.

This makes it easier to copy, paste & install all dependencies at once. In addition many users have custom setups (virtualenv, conda, .etc). Having it in one line easily allows to grap the dependencies.

* Remove 'sudo' from all pip install commands and adjust troubleshooting section.

* Make max_length and static_batch configurable.

* Fix line length.

* Fix incorrect parameters in building eval input.

* Improve comments for readability.

* Add flag to use custom training loop for keras NCF model.

* Add error check to NCF model for custom training loop + tf1.0.

* Add static batch benchmarks to estimator 

So we can distinguish how much static vs dynamic batch matter.

* change max_length for static_batch tests

* Add flag for max length

* Add a custom training loop for NCF model with TF2.0

* Fix long line in ncf_keras_main.py

* Remove dataset repeat when using custom training loop.

this is not going to help with current tf.data semantics. so removing it.

* Fixes that make transformer run.

* Remove debug print statements.

* Changed the permissions to 644.

* Fix the rest of the permissions.

* enable static batch in all benchmarks

* Restrict dist strat hack to training mode

For now we will do predict/eval without dist strat, so remove that hack in non training cases.

* Use `inputs` instead of `x` as arg name for call

Keras has different behavior based on whether the inputs are called `inputs` or not. Using `inputs` gives expected behaviors.

* Avoid extra map fn on input in dist strat case

* Update how we handle custom metrics

This new approach works with and without dist strat. The previous one didn't work with dist strat. We need to fix that but this is reasonable in meantime (b/133724664).

* Update benchmarks

* typo in metrics code

* Revert metrics change

Didn't actually work in distributed case..

250347237  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Fix linting errors in BERT benchmark test.

--
250326131  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Internal change

250315593  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Internal change

250303528  by haoyuzhang<haoyuzhang@google.com>:

    Add method docstring to fix lint error.

--

PiperOrigin-RevId: 250347237

* Run different numbers of steps on different platforms

* Add new tests for delayed performance measurement

This shuffling should help in getting shuffling each epoch.

The ".mat" files loaded in the dataset are byte files. Python 3.7 requires them to be loaded using "rb".

250009207  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Add feature in BERT to write training metrics to a summary file.

--

PiperOrigin-RevId: 250009207

249896208  by hongkuny<hongkuny@google.com>:

    Adds __init__.py

--

PiperOrigin-RevId: 249896208

Add early stopping logic to ncf keras when desired threshold is met. Also change the default batch size to match the tuned hyperparams

249883771  by hongkuny<hongkuny@google.com>:

    Creates a benchmark dir

--

PiperOrigin-RevId: 249883771

* Moved common keras code to utils.

* Initial 1 gpu benchmark

- Aligned flags with resnet example
- removed code/features that are not super useful
- eval as part of train if bleu source/ref provided
- add exp_per_second hook

* Rename benchmark classes, pass batch-size and log_steps.

* fix docstring

* Predict done with checkpoints inline

- perfzero baseclass

* steps not epochs with smoother training loop.

* do not initialize history outside loop.

* 5000 between eval not 500

* estimator to keras.

* remove epochs var.

* use range not xrange.

* 200K steps for 1 gpu

* fix global step

* Add a graph optional_next Reset benchmark.

* Fix lint error.

* Merged commit includes the following changes:
249776315  by tianlin<tianlin@google.com>:

    Internal change

249763206  by tianlin<tianlin@google.com>:

    For TF 2.0 (related to Beam Search), expand cond dims in tf.where(cond, x, y) to make all parameters broadcastable.

--
249392724  by hongkuny<hongkuny@google.com>:

    Internal change

PiperOrigin-RevId: 249776315

* Merged commit includes the following changes:
249823043  by tianlin<tianlin@google.com>:

    Bring back v2 test for predict and eval.

--

PiperOrigin-RevId: 249823043

* Add a test enabling get_next_as_optional behavior.

* Remove repeated flag.

* Remove trailing space.

* Make the name shorter.

* Fix lint error.

* Refine the benchmark name.

Adding validation every epoch allows us to view the progress during training instead of having to wait until the last eval. Mostly useful for manual runs.

Current batch size 160000 does not converge to the desired HR. So we decrease to 99k which is known to converge. Tested locally and got to 63.5 at epoch 7. Also decreasing number of epochs as I don't see any improvement after epoch 7-8.

249580533  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Internal change

249566870  by A. Unique TensorFlower<gardener@tensorflow.org>:

    Set up BERT benchmark test.

--

PiperOrigin-RevId: 249580533

* Add enable_get_next_as_optional flag.

* Set enable_get_next_as_optional to strategy.

* Add comments to explain the flag.

* Remove trailing whitespace.

* Remove trailing space.