Currently, the logs that are produced contain metadata, Apache-log style (e.g. timestamp, some latency information, etc.). They do not contain the requests (inputs) or responses (predictions). To enable these logs, see the docs:
https://cloud.google.com/ml-engine/docs/tensorflow/prediction-overview#prediction_logging
If you're interested in logging the request and responses, we're currently gathering information and would love to hear about your requirements. Please reach out to [email protected] and we'll ask you a few questions.