Open mathephysicist opened 1 year ago
I believe you are correct. Also, the associated paper makes a leap here in the described output of the TNet. The (first) TNet should output a 3x3, however the final specified layer has an output dimension of 256. Unfortunate bit of ambiguity.
https://github.com/charlesq34/pointnet/blob/539db60eb63335ae00fe0da0c8e38c791c764d2b/models/transform_nets.py#L83
You set the weights and bias to zero. Then matmul.
Hence you will always have zero updates to the TNet that is prior.