How can I get the 1.951 validation EPE with FlowNetS? (my EPE is 2.03)

[sczhou]

Hi, I trained the FlowNetS with default options, but I can not get 1.951 validation EPE, and my EPE is 2.03. Is that reasonable? And is your split-value set to 0.8?

Thanks! Shangchen

[ClementPinard]

Hello,

Default parameters are here to match original FlowNet training implementation

However, I did some hyperparameter changes to get to 1.951, have you tried the ones i used in here ? (among others, epoch size is extended to 2.5k instead of 1k)

I did not use a split value, I used the split file provided here, test set is actually much smaller than 20% of dataset, which can also explain the accuracy drop you got. (but 2.03 is actually pretty good, since it outperforms initial scores reported by FlowNetC in 2015 paper )

[sczhou]

Yeah, I tried the hyperparameters as you mentioned in your training results. Then, I'll try to use the split file instead. However, there is still a problem in making data list in flyingchairs.py. you use 'glob.iglob(os.path.join(dir,'*_flow.flo'))' to load the flo file list, but I found the iglob's return is disorder, which does not match with the FlyingChairs_train_val.txt.

Thanks for your kind reply, Regards, Shangchen

[ClementPinard]

Did not see that ! thanks for poiting it out ! A patch is incoming

[sczhou]

Ok, thanks!

[ClementPinard]

Done