I find that there maybe a bug in the ROIAlign operator. For example,
x = mx.nd.array([[[[ 0., 1., 2., 3., 4., 5.],
[ 6., 7., 8., 9., 10., 11.],
[ 12., 13., 14., 15., 16., 17.],
[ 18., 19., 20., 21., 22., 23.],
[ 24., 25., 26., 27., 28., 29.],
[ 30., 31., 32., 33., 34., 35.],
[ 36., 37., 38., 39., 40., 41.],
[ 42., 43., 44., 45., 46., 47.]]]])
// region of interest i.e. bounding box coordinates.
y = mx.nd.array([[0,0,0,4,4]])
pooled_size = (2, 2)
spatial_scale = 1.0
The result of mx.nd.ROIAlign(x, y, pooled_size, spatial_scale) is
[[[[-3.4028235e+38 -3.4028235e+38]
[-3.4028235e+38 -3.4028235e+38]]]].
When I use ROIAlign from MxNet(https://github.com/apache/incubator-mxnet/blob/master/src/operator/contrib/roi_align.cc), I get the same result as caffe2's implementation of ROIAlign(https://github.com/pytorch/pytorch/blob/master/caffe2/operators/roi_align_op.cc).
I find that there maybe a bug in the ROIAlign operator. For example,
x = mx.nd.array([[[[ 0., 1., 2., 3., 4., 5.],
[ 6., 7., 8., 9., 10., 11.],
[ 12., 13., 14., 15., 16., 17.],
[ 18., 19., 20., 21., 22., 23.],
[ 24., 25., 26., 27., 28., 29.],
[ 30., 31., 32., 33., 34., 35.],
[ 36., 37., 38., 39., 40., 41.],
[ 42., 43., 44., 45., 46., 47.]]]])
// region of interest i.e. bounding box coordinates.
y = mx.nd.array([[0,0,0,4,4]])
pooled_size = (2, 2)
spatial_scale = 1.0
The result of mx.nd.ROIAlign(x, y, pooled_size, spatial_scale) is
[[[[-3.4028235e+38 -3.4028235e+38]
[-3.4028235e+38 -3.4028235e+38]]]].
When I use ROIAlign from MxNet(https://github.com/apache/incubator-mxnet/blob/master/src/operator/contrib/roi_align.cc), I get the same result as caffe2's implementation of ROIAlign(https://github.com/pytorch/pytorch/blob/master/caffe2/operators/roi_align_op.cc).