The training EPE error explodes on FlyingChairs dataset. about flownet2-pytorch

[fitsumreda]

The Flownet2 paper actually recommends to train the networks one at a time, starting with FlowNetC, then fix this, and train FlowNetS1 and so on.
training flownet2 directly on flying-chairs would lead to the type of epe explosion you're getting.

[fitsumreda]

Could you run this command that runs flownetc, and see if you get the same error?
python main.py --batch_size 8 --model FlowNet2C --optimizer=Adam --optimizer_lr=1e-4 --loss=MultiScale --loss_norm=L1
--loss_numScales=5 --loss_startScale=4 --optimizer_lr=1e-4 --crop_size 384 512
--training_dataset FlyingChairs --training_dataset_root /path/to/flying-chairs/dataset
--validation_dataset MpiSintelClean --validation_dataset_root /path/to/mpi-sintel/clean/dataset

[PkuRainBow]

@fitsumreda But I just try to train the FlowNetC on FlyingChairs directly, I meet other problems. Could you help me?

[PkuRainBow]

It works! I am wondering whether you could provide more details, such as the well trained models by yourself.

[dvorak0]

@PkuRainBow @fitsumreda Did you get converged result with the mentioned training parameters?

Using python main.py --batch_size 8 --model FlowNet2C --optimizer=Adam --optimizer_lr=1e-4 --loss=MultiScale --loss_norm=L1 \ --validation_frequency 1 \ --loss_numScales=5 --loss_startScale=4 --optimizer_lr=1e-4 --crop_size 384 512 \ --training_dataset FlyingChairs --training_dataset_root /root/per_nas/FlyingChairs_release/data \ --validation_dataset MpiSintelClean --validation_dataset_root /root/MPI-Sintel-complete/training after 1 epoch (~3 hr in my machine), my training loss curve still keeps jumping.

[MatthewD1993]

@dvorak0 Do you get good result on training FlowNet2C? My curve is also bumping seriously while training. Very grateful if you could share your experience. Thanks!

[ericzw]

@PkuRainBow hi, have you trained flownet2 successfully by training part of flownet2 first? could you tell me how to freeze parameters?

[Queenyy]

@MatthewD1993 Hi, have you trained FlowNet2C successfully? My loss curve is also bumping while training. Thanks a lot.