Summary: 94 instances, 79 unique

Text	Count
// TODO remove the dependency on input and use instead its sizes -> save memory	2
# TODO think about a representation of batch of boxes	1
# TODO should I split them?	1
// TODO raise error if not compiled with CUDA	1
# TODO resolve this difference and make it consistent. It should be per image,	1
# TODO maybe push this to nn?	1
//TODO: ROIs come in as float, fix other blocks so they come in as same type as input.	2
//TODO: add an option for non sorted input boxes	1
# TODO make it pretty	1
// TODO: kernel implementation could stride on spatial dimension. We probably	1
# TODO: This is doubling up output	2
/* TODO: move this to a common place */	1
# TODO check if want to return list of not	1
# TODO: Remove when Batched NMS stop leading to eval metrics being all 0	1
# TODO might be better to add an extra field	2
// TODO add more checks	1
# TODO: specify init for the above	1
# TODO: Remove when Batched NMS stop leading to eval metrics being all 0	1
# TODO: remove this in the future	3
// TODO: Figure out why const-correctness doesn't work here	2
#TODO: this transpose may be needed for modularity of Detectron repo	1
# TODO: Do we need both of these?	1
// TODO: Use something less heavy duty than a big honking mutex	1
# TODO should I filter empty boxes here?	1
TO_REMOVE = 1  # TODO remove	1
# TODO: resolve this hack	2
# TODO: can this be merged into above?	1
// TODO: Stop manually allocating CUDA memory; allocate an ATen byte	1
else:  # TODO: Restore this API int32 dtype will be supported on GPUs.	1
# TODO chck if necessary	1
// TODO: Shouldn't all returned results be successful?	1
# TODO replace with get_img_info?	1
// TODO make it in a common file	3
// TODO: check that weight matches output->sizes()	2
# TODO: Test giou loss for rpn head	1
## TODO: take care of NHWC/NCHW cases for RPN-only case	1
# TODO Ideally, just remove this and let me model handle arbitrary	1
// TODO: check that output->size() matches output_sizes	2
// TODO: Consider renaming zero-indexed arguments to "self"	1
# TODO: This needs to be updated for new mask format	1
#TODO: define for CPU as well	1
# FIXME remove this once c10d fixes the bug it has	1
## TODO: NHWC kernel + transposes is faster than NCHW backward kernel	1
// TODO improve this part	1
// TODO: Add constructors for all of the descriptors	1
## TODO: we can get rid of this transpose by changing box head loss computation accordingly	1
# TODO: need to fix GN issue for NHWC	1
# FIXME remove check once we have better integration with device	1
# TODO use size instead of scale to make it robust to different sizes	1
# TODO Add losses to builders	1
# TODO this doesn't look great	1
# TODO maybe add an API for querying the ws / hs	1
# TODO: handle these two reserved arguments	1
// TODO: Go through all the checking code again and make sure	1
# TODO: move this method out of runner	2
# FIXME ideally this would be achieved with a CombinedLRScheduler,	1
TO_REMOVE = 1  # TODO remove	1
# TODO: Create a BatchedBoxlist class to keep these data structures	1
TODO: Test this	1
THCState *state = at::globalContext().lazyInitCUDA(); // TODO replace with getTHCState	1
# TODO add squeeze?	1
// TODO: Use TensorGeometry here instead of the entire Tensor, which we	1
# TODO need to make the use_07_metric format available	1
# TODO do this properly	1
//TODO: Math type is hard coded to float assuming double is not used, if needed, add a case for double as well.	2
# TODO: Use JIT to replace CUDA implementation in the future.	1
# TODO check if want to return a single BoxList or a composite	1
//TODO: larger threads-per-block might be better here, because each CTA uses 32 KB of shmem,	1
# TODO fix optimizer state dict to enable saving	1
# TODO make this nicer, this is a direct translation from C2 (but removing the inner loop)	1
# TODO check if this is the right one, as BELOW_THRESHOLD	1
# TODO remove and use only the Keypointer	1
# TODO rename x to roi_box_features, if it doesn't increase memory consumption	1
# TODO redundant, remove	1
# TODO:  Is this JIT compatible?	1
// TODO: Change other blocks producing ROI to support half type as well	2
torch.cuda.empty_cache()  # TODO check if it helps	1
# TODO:	1
# TODO maybe remove this and make it explicit in the documentation	1