search

lmb-freiburg / flownet2

FlowNet 2.0: Evolution of Optical Flow Estimation with Deep Networks
https://lmb.informatik.uni-freiburg.de/Publications/2017/IMKDB17/
Other
1k stars 318 forks source link

Reproducing Results FlowNet2 #5

Closed el3ment closed 7 years ago

el3ment commented 7 years ago

If I wanted to reproduce your results (starting from random) do I need to manually train each network independently and load their respective weights in sequence, or is there a simpler method?

nikolausmayer commented 7 years ago

Hi Robert,

from scratch, getting to the full FlowNet2 entails 5 consecutive network setups (and more steps if you count finetuning). The FlowNet2 is a fusion (1 training) of a FlowNet2-CSS-ft-sd (3 trainings, plus finetunings) and a FlowNet2-SD (1 training). It's... not a quick process.

Best, Nikolaus

el3ment commented 7 years ago

Totally understand. We are porting the model to Tensorflow and want to make sure we have a version we can train from scratch. So when I run caffe train --solver solver.prototxt on the FlowNet2 model prototxt, is it loading the weights for the FlowNet2-SD and FlowNet2-CSS-ft-sd networks? As far as I can tell from the code is that it initializes all layers using MSRA and trains it end-to-end.

Robert

eddy-ilg commented 7 years ago

Yes, this would train it end to end (its provided for fine-tuning with ChairsSDHom).

If you want to retrain the complete network stack, you have to do it step by step. We don't provide scripts for that.