You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
There has been a few mismatches between the naming conventions and input params. It would be great to standardize those, as to make code re-use much easier. To name a few:
In COCO segmentation/deteciton there is a difference between the loader expecting data_dir and root_dir.
The classes capitalization is following different standards as well COCODetectionDataset vs CoCoSegmentationDataSet.
There is coco_segmentation_dataset_params, but no coco_detection_dataset_params, but coco_detection_{network structure}_dataset_params. Change the segmentation to follow the same standard?
Similarly dataloaders.coco_segmentation_train vs dataloaders.coco2017_train_yolo_nas.
define a __str__ representation for the loaders, to avoid <torch.utils.data.dataloader.DataLoader object at 0x7fadc1d16620>
segmentation has train_loader.dataset.dataset_classes_inclusion_tuples_list, but detection doesn't. A common set of params?
edit: (a few more)
DetectionMetrics, PoseEstimationMetrics, and PreprocessSegmentationMetricsArgs
In tutorials/colabs - train_loader, train_dataloader, train_data, val_loader, valid_loader.
coco2017_rescoring_train is actually using coco2017_pose_rescoring_*
In link above PoseEstimationMetrics comment/code section error.
the metrics_to_watch is "target_IoU" for *_metrics_list that includes BinaryIoU(), but the metric is just "IoU" if it's a multiclass case with the metrics list including IoU().
Proposed Solution (Optional)
No response
The text was updated successfully, but these errors were encountered:
Hey. Thanks for the report. Indeed, the naming scheme is not consistent across the config files & models.
And in fact we already started looking into standardizing it. At this point it is hard to give an accurate estimate when exactly this will be introduced.
🚀 Feature Request
There has been a few mismatches between the naming conventions and input params. It would be great to standardize those, as to make code re-use much easier. To name a few:
data_dir
androot_dir
.COCODetectionDataset
vsCoCoSegmentationDataSet
.coco_segmentation_dataset_params
, but nococo_detection_dataset_params
, butcoco_detection_{network structure}_dataset_params
. Change the segmentation to follow the same standard?dataloaders.coco_segmentation_train
vsdataloaders.coco2017_train_yolo_nas
.__str__
representation for the loaders, to avoid<torch.utils.data.dataloader.DataLoader object at 0x7fadc1d16620>
train_loader.dataset.dataset_classes_inclusion_tuples_list
, but detection doesn't. A common set of params?edit: (a few more)
PreprocessSegmentationMetricsArgs
coco2017_rescoring_train
is actually usingcoco2017_pose_rescoring_*
Fixme
in documentationPoseEstimationMetrics
comment/code section error.metrics_to_watch
is"target_IoU"
for*_metrics_list
that includesBinaryIoU()
, but the metric is just"IoU"
if it's a multiclass case with the metrics list includingIoU()
.Proposed Solution (Optional)
No response
The text was updated successfully, but these errors were encountered: