search

Shuijing725 / CrowdNav_DSRNN

[ICRA 2021] Decentralized Structural-RNN for Robot Crowd Navigation with Deep Reinforcement Learning
https://sites.google.com/illinois.edu/crowdnav-dsrnn/home
MIT License
95 stars 23 forks source link

The Accuracy of Model Training #17

Closed AngLancer closed 1 year ago

AngLancer commented 1 year ago

Hello, I'm sorry to interrupt you. I used the parameters in /data/example_model/configs/config.py to train on my machine and got /data/dummy/checkpoints/27776.pt, and I used test.py to test my training results, the results are: TEST has success rate: 0.21, collision rate: 0.12, timeout rate: 0.67, NAV time: 20.78, total reward: 9.6464. This is far from the 90% correct rate in your /data/example_model/test/test_27776.pt.log. Why is this? My GPU is NVIDIA Tesla K80 and my CPU is three Xeon E5-2678 V3. Python 3.7, CUDA 10.2, cuDNN 7.6, Pytorch 1.6.0, Ubuntu 18.04

Shuijing725 commented 1 year ago

Can you provide the training curve and the checkpoint you tested with?

AngLancer commented 1 year ago

Sorry, I didn't save the previous results. I ran again with 2080ti, and the success rate became 90%, consistent with yours. But I still don't know why the success rate of K80 is so low. Excuse me again, I see that the success rate in /data/example_model/test/test_27776.pt.log is 90%, and the success rate of Fov360 in Figure 5 in the paper is 98%. Under what circumstances has such a high accuracy been achieved? Thank you.

Shuijing725 commented 1 year ago

Your new results look normal. If you change the random seed, change the checkpoint, or finetune the hyperparameters, you may get better results.