I think many people are participating now. This thread is mainly to share experimental observations, what strategies worked for you,improvements etc.
I used a ResNet50 with approx 10k images/class in the beginning, but it didn’t help much. Training for 1 epoch on P100 took 8 hours, and still doesn’t exceed above 0.9 on LB
Please consider discussing this on Kaggle forums, you can find the thread here. Sharing information, especially code, in a closed setting (unless everyone is part of the same team) would be breaking Kaggle rules.
2 Likes
Okay, wasn’t aware of it.
No worries, no harm done Your willingness to share your findings is much appreciated.