FlowNet-C training curve

[texastone]

I am training C using FC dataset and the long schedule, all default parameters. The flow_loss2 output is flat at 5000 from very early stage to about iteration 650000 and then starts to drop and kind of appearing normal. I was wondering if this is an expected behavior. Should I try different hyper-parameters? It seems github does not have an option to let me upload image.

[tonmoy-saikia]

Typically it drops sharply in the beginning. After 100k you should see some results which already makes sense.

[nikolausmayer]

(closed due to inactivity)