why cocat f_encoder_list[-j-2]w with the current upsampled feature ?

[no name]

      In RandLANet.py line 69, it is written:

f_decoder_i = self.decoder_blocks[j](torch.cat([f_encoder_list[-j - 2], f_interp_i], dim=1))

I'm wondering why not concatenate the f_encoder_list[-j - 1] and f_interp_i?

[qiqihaer]

      In RandLANet.py line 69, it is written:

f_decoder_i = self.decoder_blocks[j](torch.cat([f_encoder_list[-j - 2], f_interp_i], dim=1))

I'm wondering why not concatenate the f_encoder_list[-j - 1] and f_interp_i?

Take j=0 for example, the the f_interp_i is recovered from f_encoder_list[-1] by nearest_interpolation, which means f_interp_i's is related to f_encoder_list[-2].