I was fine tuning a resnet-18 net on my own dataset and got to good val results on an AWS machine.
I then copied all the fb.resnet.lua directory, including the saved checkpoints, and my dataset to a different machine. I then ran again, on this different machine, with "-resume ./checkpoint" and saw that indeed the best checkpoint is loaded - but couldn't get even near the val results I got before - it was actually like the weights were reset.
I tried comparing also the luarocks installed rocks - but all the same.
Any idea what could cause this?
Hi,
I was fine tuning a resnet-18 net on my own dataset and got to good val results on an AWS machine. I then copied all the fb.resnet.lua directory, including the saved checkpoints, and my dataset to a different machine. I then ran again, on this different machine, with "-resume ./checkpoint" and saw that indeed the best checkpoint is loaded - but couldn't get even near the val results I got before - it was actually like the weights were reset.
I tried comparing also the luarocks installed rocks - but all the same. Any idea what could cause this?
Thanks! Yotam