BasicEncoder ResidualBlock dimensions do not match with paper
christian-rauch opened this issue · comments
Christian Rauch commented
In the paper, the full (RAFT 4.8M) encoder has residual blocks with output sizes of 64 (1 x 64), 128 (2 x 64), and 192 (3 x 64), which follows the same pattern as for the smaller (RAFT-S 1M) encoder with 32 (1 x 32), 64 (2 x 32) and 96 (3 x 32).
In the code, this matches for the small encoder:
Lines 216 to 218 in aac9dd5
but for the full encoder, those values do not match:
Lines 139 to 141 in aac9dd5
Instead of [1,2,3] x 64, the pattern is here [2,3,4] x 32, so that instead of doubling the size of the feature vectors from small to full encoder, there is only a linear increase of 32.
Is this intended?