Plot of CPE during training

I’m going to implement a comparison of two models using CPE, as well as counting CPE during training in my task, so I decided to start with the Cartpole toy problem as an example. After reading the [tutorial](https://reagent.ai/usage.html#usage) and the [article](https://arxiv.org/pdf/1811.00260.pdf), I assumed that CPE's output will be stored in files during the training and can be used in Tensorboard for visualization, however, nothing like this happens. Can you tell me why?
I ran a tutorial on offline learning.
Here are the steps I've executed:
I tried to run the tutorial completely on the points:
```
1) export CONFIG=reagent/workflow/sample_configs/discrete_dqn_cartpole_offline.yaml
2)./reagent/workflow/cli.py run reagent.workflow.gym_batch_rl.offline_gym $CONFIG
3)mvn -f preprocessing/pom.xml clean package
4)rm -Rf spark-warehouse derby.log metastore_db preprocessing/spark-warehouse preprocessing/metastore_db preprocessing/derby.log
5)./reagent/workflow/cli.py run reagent.workflow.gym_batch_rl.timeline_operator $CONFIG
6)./reagent/workflow/cli.py run reagent.workflow.training.identify_and_train_network $CONFIG
7)./reagent/workflow/cli.py run reagent.workflow.gym_batch_rl.evaluate_gym $CONFIG
8) tensorboard --logdir outputs/
```
And ran into a problem in the last, 8th step: the command `tensorboard --logdir outputs/` did not output anything in the tensorboard, as if it did not find data. I assumed that there is some folder **outputs** and there are logs/data that are needed, but I did not find this folder. I've changed the command in the last step to `tensorboard --logdir .` and it helped - I was able to see the losses that were during the training, but there was no CPE either.
In order to get CPE output for Tensorboard, I've added the line `cpe_details. log_to_tensorboard()` in the file **reagent/training/dqn_trainer_base file.py** after the line 284 started the training as before, and looked at the tensorboard with the command `tensorboard --logdir .`. so I could bring CPE, but obtained only one point(as I understand it, checking CPE occurs only at the end of training on test data, that is why one point), but how to get the full schedule of values CPE? However, this is not the only problem. In addition to the fact that only one point is output, the CPE values themselves seem to be incorrect. Normalized values sometimes reach some exorbitant values, such as 200 (in _Sequential_Doubly_Robust, MAGIC, Weighted_Sequential_Doubly_Robust_), although more often they are within 10. And in the estimators _Direct_Method_Reward, Doubly_Robust_Reward, IPS_Reward_, the values are always 0.90-0.98. However, as far as I understand, the normalized value shows how many times the new agent is better than the old one that was used to create data, and a value of, for example, 10 cannot be accepted in the Cartpole task when we used a random policy to generate data.
I ran the tutorial on MAC OS 10.15.6
commit bc11359f6977b80070daaf1bb7fb725d7563551f

Could you please also elaborate on the following since it also may relate to the issues I've faced?
I found two Evaluator classes in the reagent(in the **reagent/evaluation/evaluator.py** and **reagent/ope/estimators/estimator.py**), could you explain the differences? And what is the difference between the **reagent/ope/estimators** module and the **reagent/evaluation** module?
In the [tutorial](https://reagent.ai/usage.html#usage), the command `/reagent/workflow/cli.py run reagent.workflow.gym_batch_rl.offline_gym $CONFIG` creates data from a random policy, but how to generate data from some trained one?
What probabilities should we use in calculating CPE algorithms(e.g., Importance Sampling) for DQN? logprob (that is, essentially counting probabilities after applying the greedy policy), or just take Softmax(Q-values) and count it as a probability?