search

tornadomeet / ResNet

Reproduce ResNet-v2(Identity Mappings in Deep Residual Networks) with MXNet
Apache License 2.0
556 stars 199 forks source link

resnet 260 #13

Closed mli closed 8 years ago

mli commented 8 years ago

resnet 200 is on training. any idea about 260 layers?

tornadomeet commented 8 years ago

cool! i think we should train 260 layers, and need to design the units = [,,,], @winsty say he's also consider this~

shicai commented 8 years ago

sounds crazy... how about 350 layers, or even 500 layers?

ZhongxingPeng commented 8 years ago

@tornadomeet Could you please tell me what kind of machine (hardware) you are using for training these resnets? Thanks.

austingg commented 8 years ago

@singlepetrol There are descriptions in the log directory, including GPU types, lr descreasing point. beside, the provided log file is also helpful.

shicai commented 8 years ago

any progress?

ZhongxingPeng commented 8 years ago

@austingg Thank a lot!

mli commented 8 years ago

trained several models with 200 layers using 10 machines. the final accuracy is 1% worse comparing to the one trained with single machine, maybe due to the 10x larger batch size.

i'm going to close this thread since i'm training models on the full imagenet dataset, which should be more useful for fine tune other datasets.