Gradient descent on the dilation rates, #19
Description
Hello! After reading your paper, I have gained great inspiration. However, there is one point that I don't fully understand and hope to receive a reply from you. Regarding the method mentioned in Section 3.4 of your paper, "Gradient descent on the dilation rates," I have examined your code and found that the dilation rates are hardcoded. Therefore, based on the description in your paper, can I understand it as follows: You first train the model using deformable convolutions to find the optimal parameters, and then use these parameters as the dilation rates to retrain the network?
