Total number of parameters of GA-Net-11 and GA-Net-15

[motionlife]

The total number of GC-Net is 2.9 million; (2850785 according to my TensorFlow implementation)
The total number of PSM-Net is 5.2 million; (5224768 according to my TensorFlow implementation and original author's code)

Can anyone provide me the total number of parameters of GA-Net-11 and GA-Net-15? The author doesn't seem to be consistent when reporting those numbers in the paper. They omitted the whole feature extraction layers' parameters for PSM-net and their own model. However, I found that actually there feature extraction layers are crucial to the final results, more crucial than those aggregation net in terms of final performance. So I think it's important to compare the total number of model parameters.

[18651626101]

The total number of GC-Net is 2.9 million; (2850785 according to my TensorFlow implementation) The total number of PSM-Net is 5.2 million; (5224768 according to my TensorFlow implementation and original author's code)

Can anyone provide me the total number of parameters of GA-Net-11 and GA-Net-15? The author doesn't seem to be consistent when reporting those numbers in the paper. They omitted the whole feature extraction layers' parameters for PSM-net and their own model. However, I found that actually there feature extraction layers are crucial to the final results, more crucial than those aggregation net in terms of final performance. So I think it's important to compare the total number of model parameters.

Hi, I have the same question as yours. I find that the total number of PSMNet is 5.2 million in Jia-Ren's code, however the total number of PSMNet is 3.5 million mentioned in GA-Net. Besides, when I calculate the total number of GA-Net-15, I get the different result compared with 2.3 million in the paper. I re-read the paper and I think you are right. The parameter mentioned in the paper is maybe just the 3Dconv part, ignoring the feature extraction layers.