比较与torchvision.ops.roi_align的差异

torchvision.ops.roi_align

torchvision.ops.roi_align(input: torch.Tensor, boxes: torch.Tensor, output_size: None, spatial_scale: float = 1.0, sampling_ratio: int = -1, aligned: bool = False)

更多内容详见torchvision.ops.roi_align。

mindspore.ops.ROIAlign

class mindspore.ops.ROIAlign(pooled_height, pooled_width, spatial_scale, sample_num=2, roi_end_mode=1)(features, rois)

更多内容详见mindspore.ops.ROIAlign。

差异对比

PyTorch：感兴趣区域对齐（RoI Align）。

MindSpore：感兴趣区域对齐（RoI Align）。与PyTorch相比，参数输入形式不同，对齐模式的实现也不同。

分类	子类	PyTorch	MindSpore	差异
参数	参数1	input	-	输入特征，此参数位于MindSpore算子的输入参数列表中
	参数2	boxes	-	边界框坐标，此参数位于MindSpore算子的输入参数列表中
	参数3	output_size	[pooled_height, pooled_width]	特征尺寸，MindSopre分别用2个参数表示
	参数4	spatial_scale	spatial_scale	-
	参数5	sampling_ratio	sample_num	-
	参数6	aligned	roi_end_mode	对齐的模式。PyTorch对应的参数值为False和True，MindSpore对应的参数值为0和1
输入	输入1	-	features	输入特征
	输入2	-	rois	roi坐标
输出	输出1	Tensor	Tensor	-

代码示例

# PyTorch
import numpy as np
import torch
import torchvision as tv

pooled_height, pooled_width, spatial_scale, sample_num, roi_end_mode = 3, 3, 0.25, 2, 1

features = np.array([[[[1., 2.], [3., 4.]]]]).astype(np.float32)
rois = np.array([[0, 0.2, 0.3, 0.2, 0.3]]).astype(np.float32)

features_t = torch.from_numpy(features)
rois_t = torch.from_numpy(rois)

output = tv.ops.roi_align(features_t, rois_t, (pooled_height, pooled_width), spatial_scale, sample_num, 0)
print(output)
# Out: tensor([[[[1.7000, 2.0333, 2.3667],
#                [2.3667, 2.7000, 3.0333],
#                [3.0333, 3.3667, 3.7000]]]])

# MindSpore
import mindspore as ms
from mindspore import ops

features = ms.Tensor(np.array([[[[1., 2.], [3., 4.]]]]), ms.float32)
rois = ms.Tensor(np.array([[0, 0.2, 0.3, 0.2, 0.3]]), ms.float32)

roi_align = ops.ROIAlign(pooled_height, pooled_width, spatial_scale, sample_num, 0)

output = roi_align(features, rois)
print(output)
# Out: [[[[1.7       2.0333333 2.3666668]
#         [2.3666668 2.7       3.0333335]
#         [3.0333333 3.3666668 3.7      ]]]]