Performance backbone!
phucnhs opened this issue · comments
Hi , I see you using a lot of backbone as effi,mb2,mb3,..do you compare the backbone to each other ? - do you have the performance of each one ? - i find effi quite good, why don't you use!
Thanks !
I did trained all the backbones for 20 epoch for comparison and found that EfficientNet, even B0, and ResNet34 have higher performance. However, they are much slower than ResNet18, especially EfficientNets as they use Depthwise Separable Convolution, which was not highly optimized in CUDA. So, in the limitation of time and hardware (gg colab), I only further trained ResNet18 up to 70 epoch to get the final model. The backbone can be further upgrade with stacks of BiFPN and please let me know if they achieved remarkable performance without much drop of efficiency. Tks for your response.