Training with TensorFlow on Sagemaker No module named 'tf_container'

Question

I'm trying to training TensorFlow model on AWS Sagemaker. I created container with external lib for that (Use Your Own Algorithms or Models with Amazon SageMaker).

we run a training job with TensorFlow API

from sagemaker.tensorflow import TensorFlow
estimator = TensorFlow(
  entry_point="entry.py",             # entry script
  role=role,
  framework_version="1.13.0",   
  py_version='py3',
  hyperparameters=hyperparameters,
  train_instance_count=1,                   # "The number of GPUs instances to use"
  train_instance_type=train_instance_type,
  image_name=my_image

)
estimator.fit({'train': train_s3, 'eval': eval_s3})

and got an error:

09:06:46
2019-07-23 09:06:45,463 INFO - root - running container entrypoint 
09:06:46
2019-07-23 09:06:45,463 INFO - root - starting train task 
09:06:46
2019-07-23 09:06:45,476 INFO - container_support.training - Training starting 
09:06:46
2019-07-23 09:06:45,479 ERROR - container_support.training - uncaught exception during training: No module named 'tf_container'

09:06:46
Traceback (most recent call last): File "/usr/local/lib/python3.6/dist-packages/container_support/environment.py", line 136, in load_framework return importlib.import_module('mxnet_container') File "/usr/lib/python3.6/importlib/__init__.py", line 126, in import_module return _bootstrap._gcd_import(name[level:], package, level) File "<frozen importlib._bootstrap>", line 994, in _gcd_i 
09:06:46
ModuleNotFoundError: No module named 'mxnet_container'

09:06:46
During handling of the above exception, another exception occurred:

09:06:46
Traceback (most recent call last): File "/usr/local/lib/python3.6/dist-packages/container_support/training.py", line 35, in start fw = TrainingEnvironment.load_framework() File "/usr/local/lib/python3.6/dist-packages/container_support/environment.py", line 138, in load_framework return importlib.import_module('tf_container') File "/usr/lib/python3.6/importlib/__init__.py", line 126, 
09:06:46
ModuleNotFoundError: No module named 'tf_container'

What can I do to solve this issue? how can I debug this case?

Answer 1

I'm guessing that you used your own TF container, not the SageMaker one at https://github.com/aws/sagemaker-tensorflow-container

If that's the case, your container is missing the support code needed to use the TensorFlow estimator ('tf_container' package).

The solution is to start from the SageMaker container, customize it, push it back to ECR and pass the image name to the SageMaker estimator with the 'image_name' parameter.