search

Kaixhin / Rainbow

Rainbow: Combining Improvements in Deep Reinforcement Learning
MIT License
1.56k stars 282 forks source link

Evaluate the pretrained model #77

Closed Yingdong-Hu closed 3 years ago

Yingdong-Hu commented 3 years ago

Hi, I download the pretrained V1.3 pong model and want to evaluate it. I run the following command: python main.py --game pong --model pretrained/pong.pth --evaluate But the final average reward is only -21. What's the correct method to evaluate the pretrained model ? Thank you.

Kaixhin commented 3 years ago

I "fixed" the model architecture since the 1.3 release, and it turns out that this broke all the pretrained models. You need to make a small manual change, which is now detailed in the README.