Default model_dir uses backslashes on Windows

[RobinKa]

Describe the bug
The default model_dir will contain backslashes when running the following code to start tensorflow training on Windows.

from sagemaker.tensorflow import TensorFlow

role = "<role>"

tf_estimator = TensorFlow(
    source_dir="src", script_mode=True,
    framework_version="1.15.2", py_version="py37",
    entry_point="main.py", role=role,
    train_instance_count=1, train_instance_type="ml.g4dn.xlarge",
    train_volume_size=50,
    output_path="<bucketpath>/exported_model"
)

tf_estimator.fit({
    "train": "<bucketpath>/<train dir>",
    "test": "<bucketpath>/<test dir>",
})

Trying to then upload something to the model_dir path (eg. when tensorflow tries to save checkpoints) will not work as S3 requires forward slashes. Example args that are passed when executing the above code

Args: Namespace(alpha=0.25, augment=True, batch_size=3, classes=2, crop_size=512, current_host='algo-1', epochs=3, fl_weight=0.1, gamma=2.0, hosts=['algo-1'], init_lr=0.004, layer_depth=21, mode='train', model_dir='<bucketpath>/exported_model\\tensorflow-training-2020
-07-28-14-00-00-884\\model', momentum=0.9, num_gpus=1, output_dir='output', power=0.9, regularization_scale=0.0001, sm_model_dir='/opt/ml/model', stddev=0.02, test_dir='/opt/ml/input/data/test', train_dir='/opt/ml/input/data/train')

As a workaround I just did args.model_dir = args.model_dir.replace("\\", "/") in my training code.

Expected behavior
The model_dir path should use forward slashes no matter which OS the code to initiate training is run from for S3 paths.

System information
A description of your system. Please provide:

• SageMaker Python SDK version: 1.71.1
• Framework name (eg. PyTorch) or algorithm (eg. KMeans): TensorFlow
• Framework version: 1.15.2
• Python version: 3.7.? on server, 3.7.3 locally
• CPU or GPU: GPU
• Custom Docker image (Y/N): N

[laurenyu]

Thanks for the bug report! I've opened a PR to address this: #1763

[laurenyu]

I have merged #1763 - sorry for the delay!