Fine-tuning models in Docker containers

[ZikeYan]

Hi, I've tried to use the docker to reproduce fine-tune models for SINTEL datasets as follows: enter the container as:

> $ nvidia-docker run --volume "${PWD}:/input-output:rw" -it "flownet2" /bin/bash   

enter the input-output folder(where the edited .prototxt are kept)

> $ cd .....into/the/folder

Fine-tuning the models with

> $ ./../../../flownet2/flownet2/build/tools/caffe train --solver ../solver.prototxt --weights FlowNet2_weights.caffemodel --gpu 0

It start to train, however, the fine-tuned models is much worse than the original one. in 5000 iterations the average EPE for the whole datasets is about 11, in 10000 iterations the EPE is about 9. I wonder if you could have any idea where I did is wrong.

In my opinion, the lmdb file may leads to the error, but I'm not sure. Will it matters on account that the SINTEL ground truth is .flo but not .pfm like other datasets?

Best regards

[nikolausmayer]

Hi, did you resolve this issue? Your training scores looked like they were still decreasing after 10k iterations. FLO vs PFM should not be a problem.

[ZikeYan]

Nope. I skip this problem and work on the following refinement issues for scene flow estimation. I'll work on it later and I'll close the issue if I tackle the problem afterwards. :)

[nikolausmayer]

ok! :+1:

[nikolausmayer]

closing due to inactivity.