Tensorflow script mode saves model in different directory

[harusametime]

System Information

• Framework (e.g. TensorFlow) / Algorithm (e.g. KMeans): Tensorflow
• Framework Version: 1.11.0
• Python Version: 3.6.6
• CPU or GPU: CPU
• Python SDK Version: 1.16.0
• Are you using a custom image: No

Describe the problem

When calling fit(), Tensorflow script mode does not save a trained model to s3://(default-bucket)/(default-job-name)/output/model.tar.gz, but to s3://(default-bucket)/(default-job-name)/model/ (see a picture below). However, the estimator assumes that the model is saved as model.tar.gz, estimator.deploy fails without explicitly compressing the model files and specifying the directory. I think the trained model should be saved asmodel.tar.gz in the directory output as well as other SageMaker containers do.

Minimal repro / logs

Notebook: https://gist.github.com/harusametime/ffaee7c96cfc2a50279923391dff435f
entry point: https://gist.github.com/harusametime/d0a92d715fe0610b4b8fbd11f6d42359

[harusametime]

Related to this issue? aws/sagemaker-tensorflow-training-toolkit#130

[harusametime]

Find my mistake to read model_dir. Let me check it again.

[harusametime]

Fixing a bug related to model_dir path resolves the issue. thanks.
https://gist.github.com/harusametime/d0a92d715fe0610b4b8fbd11f6d42359/revisions