When I use the official provided weights, everything worked fine. But when I use it to retrain a model on my own dataset which is built on the Vimeo90k_septuplet, things get bad. For example, if two frames are inserted between two frames, the time of the two frames appears to be distorted, and the third frame regressed from the second frame, especially in the foot of the right people.
I train it with lr=1e-4~2e-5,batch_size=16, epoch=150. Is this caused by the parameters?
When I use the official provided weights, everything worked fine. But when I use it to retrain a model on my own dataset which is built on the Vimeo90k_septuplet, things get bad. For example, if two frames are inserted between two frames, the time of the two frames appears to be distorted, and the third frame regressed from the second frame, especially in the foot of the right people.
I train it with lr=1e-4~2e-5,batch_size=16, epoch=150. Is this caused by the parameters?