I looked through this project and came across this code: examples/run_rl.py, trying it without loading an AI it seemed to work, but once I tried loading an AI already train, there were problems.
Let me explain that the time is Y for The evaluation episode is X.
Which makes little sense, but it may be due to an error on my part.
Here is my command to create the AI (without loading):
python examples/run_rl.py --log_dir experiments/test_example_rl/
And here is this one to load the old AI using its checkpoint (this command only takes a few minutes):
python examples/run_rl.py --num_episodes 100000 --log_dir experiments/test_example_rl_suite/ --load_checkpoint_path experiments/test_example_rl/checkpoint_dqn.pt --save_every 100
I would also like to point out that once the order is made, the experimental/test_example_rl_suite/ folder does not contain a checkpoint so it is impossible to reuse it because we need it to load the AI.
I would like to know if it's normal that once the AI is trained, then all new training is much faster, or if it's me who made a bad command or if I misinterpreted a code or other.
I was hesitant to use cfr to simply avoid having to save and load my agent with torch.
Hello,
I looked through this project and came across this code: examples/run_rl.py, trying it without loading an AI it seemed to work, but once I tried loading an AI already train, there were problems. Let me explain that the time is Y for The evaluation episode is X. Which makes little sense, but it may be due to an error on my part. Here is my command to create the AI (without loading):
python examples/run_rl.py --log_dir experiments/test_example_rl/And here is this one to load the old AI using its checkpoint (this command only takes a few minutes):
python examples/run_rl.py --num_episodes 100000 --log_dir experiments/test_example_rl_suite/ --load_checkpoint_path experiments/test_example_rl/checkpoint_dqn.pt --save_every 100I would also like to point out that once the order is made, the experimental/test_example_rl_suite/ folder does not contain a checkpoint so it is impossible to reuse it because we need it to load the AI. I would like to know if it's normal that once the AI is trained, then all new training is much faster, or if it's me who made a bad command or if I misinterpreted a code or other.
I was hesitant to use cfr to simply avoid having to save and load my agent with torch.
Thanks for the project!