Add shuffling to examples in example features

[johnvorsten]

Example_features are made of tensors of shape (batch_size, list_size,
feature_space). If examples are passed in a predictable order (ex.
ranked from most relevant to least relevant) the model might not generalize
well. Add 'shuffle_peritem':True to the Estimator's params to shuffle
example features and labels along list_size.

For example, when creating the estimator add the key:value pair 'shuffle_peritem':True

ranker = tf.estimator.Estimator(
    model_fn=model_fn,
    model_dir=_MODEL_DIR,
    config=run_config,
    params={'shuffle_peritem':True})

[eggie5]

Do you see any difference in practice?

[johnvorsten]

I did not see a difference in practice over my original dataset. However I’m not sure there was any relationship between my input features and my target ranking in the first place.