GitLab

PiperOrigin-RevId: 213353962

PiperOrigin-RevId: 212909744

Factor out evaluate(), continuous_eval(), and continuous_train_and_eval() from estimator_util.py into estimator_runner.py.

PiperOrigin-RevId: 212903406

PiperOrigin-RevId: 211832751

1. reshard_arrays(xs, ys): Reshards arrays in xs to match the lengths of arrays in ys.
2. uniform_cadence_light_curve(): Combines data into a single light curve with uniform cadence numbers.

PiperOrigin-RevId: 211724321

Remove the constant LONG_CADENCE_TIME_DELTA_DAYS as this is not strictly constant between light curves (although it is very close).

PiperOrigin-RevId: 211153560

This applies mainly to scrambled data (NaN time values typically come with NaN flux values, which are removed anyway, but scrambing decouples NaN time values from NaN flux values).

PiperOrigin-RevId: 209029696

Update configurations with label mapping for simulated data and new batch_size and learning_rate from vizier studies.

PiperOrigin-RevId: 208902383

PiperOrigin-RevId: 208863882

PiperOrigin-RevId: 208862798

Ensure consistent argument order, change NaN processing to remove NaN times as well, and do so before scrambling

PiperOrigin-RevId: 207804986

Modify processing pipeline to enable generation of scrambled lightcurves. Fix bugs to enable generation of inverted lightcurves.

PiperOrigin-RevId: 207595688

Inversion should be done *after* fitting a normalization curve to the stellar variability and dividing it away. Doing it before normalizing may result in unintentionally reversing the inversion. Therefore we remove the option to do it at this point.

PiperOrigin-RevId: 207309650

PiperOrigin-RevId: 205427760

PiperOrigin-RevId: 205168785

* Add --use_synthetic_data option to NCF.

* Add comment to _SYNTHETIC_BATCHES_PER_EPOCH

* Fix test

* Hopefully fix lint issue

Rollback AUTOTUNE.

Use AUTOTUNE, remove noop take, and comment fixes

Open source `Searching for Efficient Multi-Scale Architectures for Dense Image Prediction`

Open source checkpoints for mobilenetv2_dm05_coco_voc_trainaug and mobilenetv2_dm05_coco_voc_trainval

* improve default handling for eval_batch_size

* return eval_batch_size default to None

* fix syntax error

* Update resnet README with new checkpoints and SavedModels

* add more detail on channels_first vs channels_last

* fix typo

* add disclaimer about checkpoints

* set strip_default_attrs=True for SavedModel exports

* specify dtype in resnet export

* another dtype fix

* fix another dtype issue, and set --image_bytes_as_serving_input to default to False

Update installation.md

link to non-deprecated imagenet preprocessing script

Fix cvt_text citation.

Open-sourcing the checkpoint so that users could reproduce our PASCAL VOC 2012 validation set result when training on train_aug set.

* move evaluation from numpy to tensorflow

fix syntax error

don't use sigmoid to convert logits. there is too much precision loss.

WIP: add logit metrics

continue refactor of NCF evaluation

fix syntax error

fix bugs in eval loss calculation

fix eval loss reweighting

remove numpy based metric calculations

fix logging hooks

fix sigmoid to softmax bug

fix comment

catch rare PIPE error and address some PR comments

* fix metric test and address PR comments

* delint and fix python2

* fix test and address PR comments

* extend eval to TPUs

Project import generated by Copybara.

Enabling prediction in mnist_tpu.

Some changes specific to prediction. Removing traces of expected results, as this is just prediction.

Update CloudML job counts in running_pets.md

adding note only SSD models are supported