This PR proposes to update to the newer, non-beta tv_tensors API. This is necessary in order to use the latest torchvision, which only started providing aarch64 wheels in v0.17 AFAICT. However; I have no idea how datapoints is used internally e.g. in training workflows, so I rely on the experts to determine whether or not this update works comprehensively.
The package currently hard-pins to torchvision==0.15.1 to make use of the
datapointsAPI. This was removed in v0.16 in favor oftv_tensorsmoving forward. See the torchvision v0.16 release notes, tv_tensors FAQ, and tv_tensors reference for details.This PR proposes to update to the newer, non-beta
tv_tensorsAPI. This is necessary in order to use the latesttorchvision, which only started providing aarch64 wheels in v0.17 AFAICT. However; I have no idea howdatapointsis used internally e.g. in training workflows, so I rely on the experts to determine whether or not this update works comprehensively.