PR #56: Implement the logic in tpu estimator to support user provided stopping signals and save a checkpoint before stop.
Implement the logic in tpu estimator to support user provided
True/False stopping signals in the input data. When tpu estimator
encounters the user provided stopping signals with True value, it will save a
checkpoint automatically after finishing current iteration_per_loop and stops the
training. For users to use this feature, they will need add stopping signals with False
value in their real input data. And they will add some enough empty data after real input
data in the end which they know how to handle in their model code and have stopping
signals with True value.
Imported from GitHub PR https://github.com/tensorflow/estimator/pull/56
Copybara import of the project:
- ddb7cd511ae890309ea3c66553603cec903f8419 Implement the logic in tpu estimator to support and handl... by nanzhang <nanzhang@bytedance.com>
- 45b3b4b158ff8e4025b5489bbc36dfed2aa03943 Update code based on code review comments. Enable user to... by nanzhang <nanzhang@bytedance.com>
- ecdc0cd56b47bb3664a50fe3a025e098a318d675 Rename experimental_customized_tpu_infeed_outfeed_session... by nanzhang <nanzhang@bytedance.com>
- 2d2687319c96cf08aea7ae3cda5f98c327a4d9c3 Merge ecdc0cd56b47bb3664a50fe3a025e098a318d675 into 1e5c0... by ericzhangn <eric_zhn@hotmail.com>
PiperOrigin-RevId: 334224150
Change-Id: I21371581131930f7b4ba40697b6d6453482f897d
diff --git a/tensorflow/tools/api/golden/v1/tensorflow.estimator.tpu.-t-p-u-config.pbtxt b/tensorflow/tools/api/golden/v1/tensorflow.estimator.tpu.-t-p-u-config.pbtxt
index e329045..16f39f9 100644
--- a/tensorflow/tools/api/golden/v1/tensorflow.estimator.tpu.-t-p-u-config.pbtxt
+++ b/tensorflow/tools/api/golden/v1/tensorflow.estimator.tpu.-t-p-u-config.pbtxt
@@ -12,6 +12,10 @@
mtype: "<type \'property\'>"
}
member {
+ name: "experimental_feed_hook"
+ mtype: "<type \'property\'>"
+ }
+ member {
name: "experimental_host_call_every_n_steps"
mtype: "<type \'property\'>"
}
