Open edmandzhao opened 1 year ago
It has been solved. We have written a train function according to the code of RAFT and SCV. However, the performance of such training is still unclear.
Can you share a version of the training you have implemented?
Can you share a version of the training you have implemented?
Sorry, because I have modified part of the network structure here, I'm not sure whether my train code can be directly shared with you, but I can provide some help here.
First, you can take a look at the two pieces of code in his acknowledgment section. The author's overall code content is related to those two pieces. You can slightly modify the network output of the current version of the code provided by the author to facilitate the calculation of loss later. For specific train functions, you can completely refer to the train code provided in the raft and scv papers. Just make some changes on it.
The result I have reproduced is 0.05.
hello. I met some problem on training. I'm sorry to trouble you to answer this question. I can't find the relevant train file. Which file should I use to train the network.