Refactoring serving model and eval model export directory structure

Question

For Estimator-based models, models are currently not exported to the root directory, but to a deeper timestamped folder structure. Since the ModelRun artifact now contains all non model export related files, the Model artifact directory now is automatically cleaned to only contain the exported model in the base directory.

Previous Structure for Estimator-based model exports:

|-- ModelExportPath |---- EVAL_MODEL_DIR <- eval_model_dir |------ timestamped model <- eval_model_path |-------- saved_model.pb |-------- ... |---- SERVING_MODEL_DIR <- serving_model_dir |------ export |-------- exporter name |---------- timestamped model <- serving_model_path |------------ saved_model.pb |------------ ...

Refactored structure:

|-- ModelExportPath |---- EVAL_MODEL_DIR <- eval_model_dir |------ saved_model.pb |------ ... |---- SERVING_MODEL_DIR <- serving_model_dir |------ saved_model.pb |------ ...

该提问来源于开源项目：tensorflow/tfx

Answer 1

Implemented approach 2 (as discussed with Goutham). For Generic Trainer, there is no change in behavior. For Estimator based trainer, the Executor gives the user a modified set of fn_args with the serving_model_dir and eval_model_dir inside the model_run directory. Also updated examples/test cases which used keras_to_estimator.