Hello together!
The roi_align-Function in torchvision.ops only returns zeros.
Any clue why? Or where I can find help?
any help or hint would be fantastic and appreciated.
When using the whole image as a ROI I get around 4 numbers per each of the 512 roi_align-Output layer. When using real roi_coordinates its over: Just zeros.
I used spatial_scale = 0.03 (since 11/375 and 38/1242 are almost equal to 0.03).
####################################################################
####################################################################
Setup:
image [3, 375, 1242] → vgg16(feature_layers)* → roi_align** = y = [n x 512 x 11 x 38] = 0.
Where n is the index of the roi.
*vgg16:
model = models.vgg16(weights='IMAGENET1K_V1')
## remove the classifier completly
model.classifier = nn.Sequential(*list(model.classifier.children())[:-7])
## remove the avgpooling
model.avgpool = nn.Sequential(*list(model.avgpool.children())[:-1])
→ Output shape: [1, 512, 11, 38] (but flattend
I do not resize the input images for the vgg16.
**roi_align
output_size2 = (11, 38)
spatial_scale = 0.03
rois_test = torch.FloatTensor([
[0.0000, 1056.1600, 190.5000, 1241.0000, 374.0000],
[0.0000, 359.4300, 179.3000, 516.3000, 270.9700],
[0.0000, 600.1900, 163.5600, 633.5700, 188.6000]])
# Test; take whole image as one ROI
rois_whole_image = torch.FloatTensor([ [ 0, 0, 0, 1242, 375] ])
rois_x = rois.cuda()
z = img_feature.to("cuda").view(-1,512,11,38) # since the Tensor is flatten; reshape it.
y = torchvision.ops.roi_pool(z, rois_x, output_size2)
`