search

dr-costas / dnd-sed

Sound event detection with depthwise separable and dilated convolutions.
https://arxiv.org/abs/2002.00476
Other
52 stars 8 forks source link

No sigmoid on logits? #1

Closed shuki-hpcnt closed 4 years ago

shuki-hpcnt commented 4 years ago

It seems like that metric functions are using direct logit values from Linear and calculate positive/negative by thresholding with 0.5, but I cannot find any sigmoid or similar variants that transforms logits into probability estimations. Am I missing something?

dr-costas commented 4 years ago

Hi,

No, you are absolutely right, this is a bug. The thing is that this is not the only one. For example, there is no f1_func argument at the call of training function, file experiments/_processes.py, line 287.

The bugs are caused because the repository is a refactored version of the code used in the experiments, mainly cleaned and structured in a more friendly way for other people. The purpose of the initial form of repo was to showcase to the reviewers the code and prove that the code will be released publicly.

I'm planning to fix the known bugs by this Monday.

I'm sorry for that. I will let you know (by posting in this issue) when I push the fixes.

shuki-hpcnt commented 4 years ago

Thanks for the reply! Looking forward to the fixes. BTW, the other bug that you mentioned is not the bug since there is **common_kwargs which includes f1_func argument. I actually cloned the code and succeeded to run the experiment :D

dr-costas commented 4 years ago

BTW, the other bug that you mentioned is not the bug since there is **common_kwargs which includes f1_func argument. I actually cloned the code and succeeded to run the experiment :D

I must have fixed it already, good to know! :D

dr-costas commented 4 years ago

Fixed. Thank you for spotting that.

I will close this issue, please feel free to make a new issue if you spot something else.