output roi all zeros

[qchenclaire]

trafficstars

Hi I tried to test ROIAlign on images with rois = roi_align(detections, boxes, box_index) detections' shape is torch.Size([1, 3, 271, 271]) and boxes looks like tensor([[ 151.6779, 18.8237, 270.0000, 84.2876], [ 175.6971, 9.2199, 255.9987, 92.7847], [ 165.4188, 0.0000, 233.8400, 119.7061], [ 134.8676, 25.9375, 270.0000, 79.1737]], device='cuda:0') and box_index looks like tensor([ 0, 0, 0, 0], dtype=torch.int8, device='cuda:0') The output shape is (4, 50, 50, 3), which means 4 cropped images. But I got only the first image that looks correct and the rest 3 are all zeros.

[longcw]

The box index should be IntTensor but not int8

Qi (Claire) Chen [email protected] 于 2019年3月12日周二 上午6:53写道：

Hi I tried to test ROIAlign on images with rois = roi_align(detections, boxes, box_index) detections' shape is torch.Size([1, 3, 271, 271]) and boxes looks like tensor([[ 151.6779, 18.8237, 270.0000, 84.2876], [ 175.6971, 9.2199, 255.9987, 92.7847], [ 165.4188, 0.0000, 233.8400, 119.7061], [ 134.8676, 25.9375, 270.0000, 79.1737]], device='cuda:0') and box_index looks like tensor([ 0, 0, 0, 0], dtype=torch.int8, device='cuda:0') The output shape is (4, 50, 50, 3), which means 4 cropped images. But I got only the first image that looks correct and the rest 3 are all zeros. [image: 0] https://user-images.githubusercontent.com/24658370/54163296-e86aab80-442e-11e9-9750-1a52b7338a17.jpg [image: 1] https://user-images.githubusercontent.com/24658370/54163297-e86aab80-442e-11e9-9abf-36d74aab23fc.jpg [image: 2] https://user-images.githubusercontent.com/24658370/54163298-e86aab80-442e-11e9-8d01-bf3d05136816.jpg [image: 3] https://user-images.githubusercontent.com/24658370/54163299-e86aab80-442e-11e9-91a2-c23f3c540022.jpg

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/longcw/RoIAlign.pytorch/issues/25, or mute the thread https://github.com/notifications/unsubscribe-auth/AF6UgJr-CGb8Oaqx4q6-12nSPxDwuj9sks5vVt5lgaJpZM4bpuTJ .

[qchenclaire]

The box index should be IntTensor but not int8 Qi (Claire) Chen [email protected] 于 2019年3月12日周二 上午6:53写道： … Hi I tried to test ROIAlign on images with rois = roi_align(detections, boxes, box_index) detections' shape is torch.Size([1, 3, 271, 271]) and boxes looks like tensor([[ 151.6779, 18.8237, 270.0000, 84.2876], [ 175.6971, 9.2199, 255.9987, 92.7847], [ 165.4188, 0.0000, 233.8400, 119.7061], [ 134.8676, 25.9375, 270.0000, 79.1737]], device='cuda:0') and box_index looks like tensor([ 0, 0, 0, 0], dtype=torch.int8, device='cuda:0') The output shape is (4, 50, 50, 3), which means 4 cropped images. But I got only the first image that looks correct and the rest 3 are all zeros. [image: 0] https://user-images.githubusercontent.com/24658370/54163296-e86aab80-442e-11e9-9750-1a52b7338a17.jpg [image: 1] https://user-images.githubusercontent.com/24658370/54163297-e86aab80-442e-11e9-9abf-36d74aab23fc.jpg [image: 2] https://user-images.githubusercontent.com/24658370/54163298-e86aab80-442e-11e9-8d01-bf3d05136816.jpg [image: 3] https://user-images.githubusercontent.com/24658370/54163299-e86aab80-442e-11e9-91a2-c23f3c540022.jpg — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#25>, or mute the thread https://github.com/notifications/unsubscribe-auth/AF6UgJr-CGb8Oaqx4q6-12nSPxDwuj9sks5vVt5lgaJpZM4bpuTJ .

After I changed to torch.int, that works! Thanks!

[zxyyxzz]

The box index should be IntTensor but not int8 Qi (Claire) Chen [email protected] 于 2019年3月12日周二 上午6:53写道： … Hi I tried to test ROIAlign on images with rois = roi_align(detections, boxes, box_index) detections' shape is torch.Size([1, 3, 271, 271]) and boxes looks like tensor([[ 151.6779, 18.8237, 270.0000, 84.2876], [ 175.6971, 9.2199, 255.9987, 92.7847], [ 165.4188, 0.0000, 233.8400, 119.7061], [ 134.8676, 25.9375, 270.0000, 79.1737]], device='cuda:0') and box_index looks like tensor([ 0, 0, 0, 0], dtype=torch.int8, device='cuda:0') The output shape is (4, 50, 50, 3), which means 4 cropped images. But I got only the first image that looks correct and the rest 3 are all zeros. [image: 0] https://user-images.githubusercontent.com/24658370/54163296-e86aab80-442e-11e9-9750-1a52b7338a17.jpg [image: 1] https://user-images.githubusercontent.com/24658370/54163297-e86aab80-442e-11e9-9abf-36d74aab23fc.jpg [image: 2] https://user-images.githubusercontent.com/24658370/54163298-e86aab80-442e-11e9-8d01-bf3d05136816.jpg [image: 3] https://user-images.githubusercontent.com/24658370/54163299-e86aab80-442e-11e9-91a2-c23f3c540022.jpg — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#25>, or mute the thread https://github.com/notifications/unsubscribe-auth/AF6UgJr-CGb8Oaqx4q6-12nSPxDwuj9sks5vVt5lgaJpZM4bpuTJ .

Hi : hello, i am using roi_align to do some works, but i also meet the problem: the code is below: `
import torch from torchvision.ops import RoIAlign, RoIPool, roi_align import numpy as np

output_size = (7,7)
spatial_scale = 1/4
sampling_ratio = 2

x = torch.zeros((1,1,117,117), dtype=torch.float)
x[:,:,53:89,88:102]=100.0

rois = torch.tensor([
    # [0,0.0,0.0,0.0,0.0],
    [0.0, 53.0, 89.0, 88.0, 102.0],
])
count = torch.tensor(0).view(1)


ya = roi_align(x, rois, output_size, sampling_ratio=1)
print(ya)

`

the out put is: tensor([[[[0., 0., 0., 0., 0., 0., 0.], [0., 0., 0., 0., 0., 0., 0.], [0., 0., 0., 0., 0., 0., 0.], [0., 0., 0., 0., 0., 0., 0.], [0., 0., 0., 0., 0., 0., 0.], [0., 0., 0., 0., 0., 0., 0.], [0., 0., 0., 0., 0., 0., 0.]]]])

can u help me?

[cswaynecool]

I met the same problem, this happens when more than half of the video memory is used.