Kichang Kim
Kichang Kim
Hi. I originally implemented DeepDanbooru by using Microsoft's CNTK. With TensorFlow, there are some different behaviours to CNTK. So I testd some parameter changes for TensorFlow: 1. Increased learning rate:...
> 1girl, red_hair, black_hair,...1000 other tags...,blah 1,1,0,...,0 Oh, it looks fine. Additionally, I filtered training dataset, using only images which has 20 or more general tags. Images which has <...
Yes, simple multi-label classification has many limitations for semantic recognition. That is why I removed "copyright tags" from training data.
Thanks for reporting. I found that latest model has some issues when estimating repeated patterns (like checker, polka dots). I will change loss function to focal loss for the next...
I don't have any benchmark test/score of DeepDanbooru for latest model.
I implemented and tested focal-loss, but it makes training speed be too slow. With same epoch (30), it still has too high loss and low F1 score. I think that...
> I would like to know what your learning rate was? still 5? Yes, it is still 5.
I think that about 2.5~3 million images are used. (general tags >= 20)
I agree that current API is quite incomplete now. I'll improve/rewrite all of APIs in the next version. (no ETA yet)
Same issue here. With latest ACRA, Acralyzer, CentOS 6.5 on couchdb 1.6 built fron source. All of reports has NULL uptime and empty USER_CRASH_DATE and empty timestamp.