pytorch/cuda_utils/smcv_utils/cuda/nhwc/conv.cpp (10 lines):
	- line 48: // TODO: Go through all the checking code again and make sure
	- line 210: // TODO: check that output->size() matches output_sizes
	- line 211: // TODO: check that weight matches output->sizes()
	- line 232: // TODO: check that output->size() matches output_sizes
	- line 233: // TODO: check that weight matches output->sizes()
	- line 256: // TODO: Use TensorGeometry here instead of the entire Tensor, which we
	- line 307: // TODO: Use something less heavy duty than a big honking mutex
	- line 333: // TODO: Stop manually allocating CUDA memory; allocate an ATen byte
	- line 436: // TODO: Shouldn't all returned results be successful?
	- line 788: // TODO: Consider renaming zero-indexed arguments to "self"


pytorch/cuda_utils/smcv_utils/cuda/ROIAlign_cuda.cu (8 lines):
	- line 9: // TODO make it in a common file
	- line 65: // TODO: Change other blocks producing ROI to support half type as well
	- line 176: // TODO: Change other blocks producing ROI to support half type as well
	- line 484: //TODO: Math type is hard coded to float assuming double is not used, if needed, add a case for double as well.
	- line 486: //TODO: ROIs come in as float, fix other blocks so they come in as same type as input.
	- line 523: // TODO remove the dependency on input and use instead its sizes -> save memory
	- line 563: //TODO: Math type is hard coded to float assuming double is not used, if needed, add a case for double as well.
	- line 565: //TODO: ROIs come in as float, fix other blocks so they come in as same type as input.


pytorch/sagemakercv/core/structures/keypoint.py (4 lines):
	- line 10: # FIXME remove check once we have better integration with device
	- line 18: # TODO should I split them?
	- line 129: # TODO this doesn't look great
	- line 153: # TODO make this nicer, this is a direct translation from C2 (but removing the inner loop)


pytorch/sagemakercv/data/datasets/evaluation/coco/coco_eval.py (3 lines):
	- line 237: # TODO replace with get_img_info?
	- line 303: # TODO maybe remove this and make it explicit in the documentation
	- line 444: # TODO make it pretty


pytorch/cuda_utils/smcv_utils/cuda/nhwc/Descriptors.h (3 lines):
	- line 36: // TODO: Add constructors for all of the descriptors
	- line 91: // TODO: Figure out why const-correctness doesn't work here
	- line 233: // TODO: Figure out why const-correctness doesn't work here


pytorch/cuda_utils/smcv_utils/cuda/ROIPool_cuda.cu (3 lines):
	- line 10: // TODO make it in a common file
	- line 155: // TODO remove the dependency on input and use instead its sizes -> save memory
	- line 169: // TODO add more checks


pytorch/sagemakercv/layers/iou_loss.py (3 lines):
	- line 186: # TODO: remove this in the future
	- line 224: # TODO: remove this in the future
	- line 261: # TODO: remove this in the future


pytorch/sagemakercv/core/structures/boxlist_ops.py (3 lines):
	- line 44: # TODO maybe add an API for querying the ws / hs
	- line 110: #TODO: define for CPU as well
	- line 113: # TODO redundant, remove


pytorch/sagemakercv/detection/roi_heads/mask_head/inference.py (2 lines):
	- line 11: # TODO check if want to return a single BoxList or a composite
	- line 193: # TODO:  Is this JIT compatible?


tensorflow/sagemakercv/utils/runner/runner.py (2 lines):
	- line 183: # TODO: move this method out of runner
	- line 211: # TODO: This is doubling up output


pytorch/cuda_utils/smcv_utils/cuda/nms.cu (2 lines):
	- line 83: THCState *state = at::globalContext().lazyInitCUDA(); // TODO replace with getTHCState
	- line 126: // TODO improve this part


pytorch/sagemakercv/utils/runner/runner.py (2 lines):
	- line 125: # TODO: move this method out of runner
	- line 148: # TODO: This is doubling up output


tensorflow/sagemakercv/core/box_utils.py (2 lines):
	- line 402: # TODO: can this be merged into above?
	- line 488: # TODO: Do we need both of these?


pytorch/sagemakercv/core/structures/segmentation_mask.py (2 lines):
	- line 98: # TODO chck if necessary
	- line 137: # TODO add squeeze?


pytorch/sagemakercv/data/datasets/coco.py (2 lines):
	- line 94: # TODO might be better to add an extra field
	- line 197: # TODO might be better to add an extra field


pytorch/sagemakercv/detection/roi_heads/keypoint_head/inference.py (2 lines):
	- line 36: # TODO remove and use only the Keypointer
	- line 107: # TODO do this properly


pytorch/sagemakercv/detection/rpn/anchor_generator.py (2 lines):
	- line 165: # TODO check if want to return list of not
	- line 186: # TODO: Create a BatchedBoxlist class to keep these data structures


pytorch/sagemakercv/core/box_coder.py (2 lines):
	- line 37: TO_REMOVE = 1  # TODO remove
	- line 69: TO_REMOVE = 1  # TODO remove


tensorflow/sagemakercv/core/roi_ops.py (2 lines):
	- line 32: # TODO: Remove when Batched NMS stop leading to eval metrics being all 0
	- line 292: # TODO: Remove when Batched NMS stop leading to eval metrics being all 0


pytorch/sagemakercv/core/structures/image_list.py (1 line):
	- line 72: # TODO Ideally, just remove this and let me model handle arbitrary


pytorch/sagemakercv/data/datasets/evaluation/voc/voc_eval.py (1 line):
	- line 13: # TODO need to make the use_07_metric format available


pytorch/sagemakercv/layers/roi_align.py (1 line):
	- line 39: ## TODO: NHWC kernel + transposes is faster than NCHW backward kernel


pytorch/cuda_utils/smcv_utils/nms.h (1 line):
	- line 33: // TODO raise error if not compiled with CUDA


tensorflow/sagemakercv/training/losses/rpn_losses.py (1 line):
	- line 72: # TODO: Test giou loss for rpn head


pytorch/sagemakercv/layers/sigmoid_focal_loss.py (1 line):
	- line 8: # TODO: Use JIT to replace CUDA implementation in the future.


pytorch/sagemakercv/detection/roi_heads/mask_head/roi_mask_predictors.py (1 line):
	- line 42: #TODO: this transpose may be needed for modularity of Detectron repo


pytorch/sagemakercv/detection/detector/generalized_rcnn.py (1 line):
	- line 93: ## TODO: take care of NHWC/NCHW cases for RPN-only case


pytorch/cuda_utils/smcv_utils/cuda/nhwc/UpSampleNearest2d.cu (1 line):
	- line 172: // TODO: kernel implementation could stride on spatial dimension. We probably


pytorch/sagemakercv/detection/roi_heads/keypoint_head/loss.py (1 line):
	- line 93: # TODO check if this is the right one, as BELOW_THRESHOLD


pytorch/sagemakercv/detection/backbone/fpn.py (1 line):
	- line 68: # TODO use size instead of scale to make it robust to different sizes


pytorch/sagemakercv/utils/runner/hooks/logger/text.py (1 line):
	- line 37: # TODO: resolve this hack


pytorch/sagemakercv/detection/roi_heads/box_head/roi_box_feature_extractors.py (1 line):
	- line 85: ## TODO: we can get rid of this transpose by changing box head loss computation accordingly


tensorflow/sagemakercv/training/trainers.py (1 line):
	- line 46: # TODO:


tensorflow/sagemakercv/core/spatial_transform_ops.py (1 line):
	- line 196: else:  # TODO: Restore this API int32 dtype will be supported on GPUs.


pytorch/sagemakercv/layers/ce_loss.py (1 line):
	- line 118: # TODO: handle these two reserved arguments


pytorch/sagemakercv/detection/roi_heads/box_head/inference.py (1 line):
	- line 59: # TODO think about a representation of batch of boxes


pytorch/sagemakercv/utils/comm.py (1 line):
	- line 137: TODO: Test this


pytorch/sagemakercv/layers/make_layers.py (1 line):
	- line 117: # TODO: need to fix GN issue for NHWC


pytorch/sagemakercv/training/optimizers/schedulers/lr_scheduler.py (1 line):
	- line 8: # FIXME ideally this would be achieved with a CombinedLRScheduler,


pytorch/sagemakercv/core/structures/bounding_box.py (1 line):
	- line 181: # TODO should I filter empty boxes here?


tensorflow/sagemakercv/data/coco/dataloader_utils.py (1 line):
	- line 350: # TODO: This needs to be updated for new mask format


pytorch/sagemakercv/detection/roi_heads/roi_heads.py (1 line):
	- line 25: # TODO rename x to roi_box_features, if it doesn't increase memory consumption


pytorch/sagemakercv/inference/tester.py (1 line):
	- line 20: torch.cuda.empty_cache()  # TODO check if it helps


pytorch/cuda_utils/smcv_utils/cuda/nms_batched.cu (1 line):
	- line 147: //TODO: add an option for non sorted input boxes


pytorch/cuda_utils/smcv_utils/cuda/nhwc/UpSample.cuh (1 line):
	- line 10: /* TODO: move this to a common place */


pytorch/sagemakercv/layers/smooth_l1_loss.py (1 line):
	- line 5: # TODO maybe push this to nn?


pytorch/sagemakercv/data/samplers/distributed.py (1 line):
	- line 3: # FIXME remove this once c10d fixes the bug it has


pytorch/sagemakercv/detection/backbone/resnet.py (1 line):
	- line 291: # TODO: specify init for the above


pytorch/sagemakercv/utils/runner/hooks/checkpoint.py (1 line):
	- line 45: # TODO fix optimizer state dict to enable saving


pytorch/sagemakercv/detection/rpn/inference.py (1 line):
	- line 400: # TODO resolve this difference and make it consistent. It should be per image,


tensorflow/sagemakercv/training/builder.py (1 line):
	- line 27: # TODO Add losses to builders


tensorflow/sagemakercv/utils/runner/hooks/logger/text.py (1 line):
	- line 53: # TODO: resolve this hack


pytorch/cuda_utils/smcv_utils/cuda/generate_mask_targets.cu (1 line):
	- line 439: //TODO: larger threads-per-block might be better here, because each CTA uses 32 KB of shmem,


pytorch/cuda_utils/smcv_utils/cuda/SigmoidFocalLoss_cuda.cu (1 line):
	- line 14: // TODO make it in a common file