Hi, I have been using CineMA for downstream tasks (regression on heart phenotypes) and discovered a discrepancy in how 3D volumes are handled compared to standard PyTorch Conv3d expectations.
The model appears to have been trained with the spatial dimensions ordered as (Batch, Channel, Height, Width, Depth) instead of the PyTorch standard (Batch, Channel, Depth, Height, Width). I suspect that feeding this raw data directly into the CineMA encoder results in Conv3d treating Height as the Depth axis.
Could you perhaps verify if this was intentional or accidental?
Thanks!
Hi, I have been using CineMA for downstream tasks (regression on heart phenotypes) and discovered a discrepancy in how 3D volumes are handled compared to standard PyTorch Conv3d expectations.
The model appears to have been trained with the spatial dimensions ordered as (Batch, Channel, Height, Width, Depth) instead of the PyTorch standard (Batch, Channel, Depth, Height, Width). I suspect that feeding this raw data directly into the CineMA encoder results in Conv3d treating Height as the Depth axis.
Could you perhaps verify if this was intentional or accidental?
Thanks!