google-deepmind
ResNet implementation for TAPIR different from typical
Why has TAPIR used a different implementation for the ResNet than the typical one used (https://github.com/pytorch/vision/blob/main/torchvision/models/resnet.py, particularly BasicBlock vs BlockV2)?
Hi @swarnim-j
Our ResNet in TAPIR implementation was impacted by TSM-ResNet in the TAP-Net project and ResNet in RAFT and PIPs. We also did a little internal exploration on the ResNet structure to try to find a slightly better backbone (i.e. replacing batchnorm with instancenorm). In the end it differs from the original resnet slightly.
Performance wise, it probably won't change that much if you directly replace the resnet backbone with the original one and train. We have never tried, would be interesting to know how much gap in the end.
