predicted HOG target visualization

[ghost]

Hey, I am grateful for your excellent work . As mentioned in your paper and maskfeat, do you know how we could visualize the predicted HOG target? During model training, I set the parameter gaussian_window in the HOGLayerC class to 16 and args.norm_ fix_ loss to false. By reshaping the HOG target[B,64,108], I can get the general outline of HOG target feature, but it is still different from the one in the paper.