Failed to load the trained model using kubeflow deployer component

[jinchihe]

Used kubeflow deployer pipeline component to deploy some trained mode on kubeflow cluster, but the failed to load trained model, logs are as following:

2019-04-08 13:25:08.014242: I external/org_tensorflow/tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2019-04-08 13:25:08.217134: I external/org_tensorflow/tensorflow/cc/saved_model/loader.cc:291] SavedModel load for tags { serve }; Status: fail. Took 408319 microseconds.
2019-04-08 13:25:08.222164: E tensorflow_serving/util/retrier.cc:38] Loading servable: {name: mymodel version: 3} failed: Not found: Op type not registered 'NonMaxSuppressionV3' in binary running on highway-v1-5b4d556cc9-62j9p. Make sure the Op and Kernel are registered in the binary running in this process.

[jinchihe]

Checked details, seems that's caused by the tf-serving version is too old, used kubeflow v0.2.5 in Dockerfile. After upgrade tf-serving, the model can be loaded successfully.

The TF-Serving has been upgraded to 1.11.1 from kubeflow 0.4.0. We should upgrade to use new tf-serving in kubeflow in Dockerfile.

Will send a PR for the issue.