search

feiyuhuahuo / Anomaly_Prediction

Pytorch implementation of anomaly prediction.
78 stars 22 forks source link

About FlowNet and optimizer #9

Closed LanduRi closed 4 years ago

LanduRi commented 4 years ago

Hello, I have seen that you have provided two kinds of flownet options. I would like to ask which performance is better. I tested it. It seems that Lite flownet and flownet2sd have similar performance. Is that right?

Another question, I read your commits history. You tried SGD optimizer, and then you changed back to Adam. Is Adam better in your experiment?

feiyuhuahuo commented 4 years ago

When I was visualizing these two flownets, their performance seems to be close. Both flownets were not trained on the dataset, and I don't know how to train them. So I used them directly. The paper said Lite flownet has better performance. Adam can get better performance than SGD, but not too much, at most 2%, I can't remember clearly.

LanduRi commented 4 years ago

Thank you very much for your reply.

Have you tried the segmented learning rate as the author did? LRATE_G = [0.0002, 0.00002] LRATE_G_BOUNDARIES = [50000]

feiyuhuahuo commented 4 years ago

I forgot it. But learning rate is not crucial if it is not changed too much.