Sagemaker rlestimator - deployment endpoint giving different prediction than evaluation

[troybvo]

I ran into the problem of using reinforcement learning dueling double deep q network on sagemaker. Basically the endpoint predictions almost always predicts 1 values, while the same dataset in evaluation (same model checkpoint) gives even predictions among the classes. I used this operation for the output node main_level/agent/main/online/network_0/dueling_q_values_head_0/output, which is the last operation in a dueling q network.

Is this a bug within the checkpoint deployment? or my output operation is wrong? Any help would be appreciated.

[martinRenou]

Were you able to find your solution? Could you give more information about your issue and (if applicable) your solution?

Otherwise we may close this issue for triaging.

[mufaddal-rohawala]

Thank you for opening this issue. Closing this issue as per the above comment. Please feel free to reopen if you continue to see this issue with the latest sagemaker version.