estimator

estimator copied to clipboard

bugfix of export_savedmodel

1

Here need to do a conversion, otherwise, the user may change the input_receiver.features that will cause the input_receiver.features to be changed.

wu-yy

Remove first export in BestExporter to avoid exporting worse model.

4

First export mechanism in BestExporter will export worse model when trainig tasks restart. @tensorflower-gardener This bug is attached with this [issue](https://github.com/tensorflow/tensorflow/issues/47396). Please review it and give some suggestions. Thanks.

wuxianxingkong

Minimal support for SavedModel in WarmStartSettings

3

This PR adds support for using a SavedModel with `WarmStartSettings` by passing `[saved_model_export_dir]/variables` to `WarmStartSettings.ckpt_to_initialize_from` as discussed in https://github.com/tensorflow/tensorflow/issues/25400. It may be nicer to just change `WarmStartSettings` to accept the...

dwyatte

Best exporter support final best evaluation result output

5

### Why In estimator, we need to print final best eval result when using the best exporter.

zuston

_save_first_checkpoint in tf.keras.estimator.model_to_estimator should only be called on chief

6

when run in distribute mode, it will cause race condition of `_save_first_checkpoint`. So, we should judge if its task_type is chief.

sharkdtu

fix eval hooks with distribute strategy

2

Cherrypicking fix for GitHub Issue https://github.com/tensorflow/tensorflow/issues/28018 to Estimator 1.15 branch. PiperOrigin-RevId: 314226004

sshrdp

Warn fix optimizers warning

1

Replace the deprecated `inspect.getargspec` with `inspect.getfullargspec` as the first is deprecated in favor of the latter. See here: https://docs.python.org/3/library/inspect.html#inspect.getargspec

CloseChoice

copy of input_receiver.features

2

deepcopy of input_receiver.features avoid changing input_receiver.features in model_fn

wu-yy