Confused about feat split and your training settings

[williamhyin]

This is your temporal fusion parameters:

temporal_fusion=dict( type='TransformAwareGRU', grid_config=bev_grid_config, transform_aware_config=dict( in_channels=C_img * 2, out_channels=C_img * 2, transform_channels=12, kernel_size=1, padding=0,), encoder_config=dict(curr_channels=C_img, hidden_channels=C_img*2, kernel_size=5), used_prev_frames=1, empty_padding='zero', prev_bev_channels=C_img*2),

question:

1. why transform_aware_config.in_channels = C_img * 2
2. why encoder_config(gru) curr_channels+hidden_channels =C_img*3？
   i see that prev, curr = feat.split([self.hidden_channels, self.curr_channels], dim=1), you only have 1 history frame(used_prev_frames=1), how to plit to 160 +80 ?

please help to solve my Confuse.