I set n_input_scans to 8 and retrained salsanext. However, the score on the test set was only 0.58
yangpan22223 opened this issue · comments
Hello, I set n_input_scans to 8 and retrained salsanext. However, the score on the test set was only 0.58. What caused this? Can you provide your training parameter settings or log for reference?
@yangpan22223, thanks for using our code.
Several things you could check:
- The final result reported in the paper was N=8 and enhanced with semantics. You may check the results after fusing the semantics and a simple way is given here
- During inferring whether KNN is enabled in the config file.
- Whether you are using the latest version 1.1, where shuffle and transform are now set to default as True, which will influence the performance.
- Which graphic and batch_size you are using. Different batch_size will also influence the performance.
All the training parameter settings and logs have already been provided in the collection of downloads.
I hope you have fun with our LiDAR-MOS benchmark and code ;-)
Thanks for the answer.
I have checked my setting:
- My result was N=8 and enhanced with semantics
- KNN is enabled when inferring.
- I used the latest version 1.1
- I use pytorch framework for training and batch_size=24.
I have a question, are there any pre-trained models that need to be loaded when training salsanext? My salsanext is trained from scratch.
There is no pre-trained step. To verify the settings, could you please train it again with 150 epochs?
The randomized initial weights of the network may also influence the results a little bit, but there are multiple users already reproduced similar results as reported in the paper.
What is your account on the benchmark? I could also help you to check the results.
Since there is no update for a while, I would like to close this issue. Feel free to reopen it, if needed.