pytorch 1.2.0 c++ interface (new-style torch.autograd.Function)

[meikuam]

Get rid of deprecation warnings with old-style autograd.Function.

[xiaoyaoyang]

Tried to install it with OSX Catalina, and got error error: unsupported option '-fopenmp'. Do you think we should replace this roi_align with roi_align from torchvision?

[meikuam]

For '-fopenmp' you could just remove this flag from setup.py file (adds OpenMP capability to CPU implementation): https://github.com/meikuam/RoIAlign.pytorch/blob/master/setup.py#L9 https://github.com/meikuam/RoIAlign.pytorch/blob/master/setup.py#L19

I use this operation for backward compatibility of my project. Since torchvision 0.3 and pytorch 1.1.0 you can use roi_align from: torchvision.ops.roi_align But there are some differences in arguments. Like for this implementation you should pass three args:

• featuremap - input tensor with shape (NxCxHxW);
• boxes - float tensor (Mx4) with boxes with (x1, y1, x2, y2) without normalization (coords are corresponded to featuremap HxW);
• box_ind - int tensor (M) with batch indexes of bboxes.

torchvision implementation reqiures:

• input - input tensor with shape (NxCxHxW);
• boxes - (Tensor[K, 5] or List[Tensor[L, 4]]): the box coordinates in first case: tensor with elements (batch_index, x1, y1, x2, y2), second case: each list element contains bboxes for corresponded batch index;
• output_size - crop size (h, w);
• spatial_scale - a scaling factor for bbox coordinates;
• sampling_ratio - number of sampling points in the interpolation grid.

I think if you intend to use torchvision.models.detection.maskrcnn_resnet50_fpn, your choise is using torchvision implementation.

I did not make comparisons on the speed of work, but I think that they are similar.