PointRend: no 'pred_masks' field in instances when there are no predictions

This commit changes behavior of PointRendMaskHead when there are no predictions on image. It does not add field pred_masks as before this commit. I expect this is not wanted behavior, as MaskRCNN also adds this field.

I think it’s because there’s no call mask_rcnn_inference now, when there are no predictions. before now

I can provide MWE if needed.

[ppwwyyxx]

It would be good to make sure all components work with empty inputs - they are so common in detection models. For the case in implicit pointrend, replacing 1x1 conv with a matmul will support empty size and also reads more clear to me:

import torch
from torch.nn import functional as F

num_instances = 0
C = 8
Cout = 16
num_points = 10

features = torch.rand(1, num_instances * C, num_points)
weights = torch.rand(num_instances * Cout, C, 1)
bias = torch.rand(num_instances * Cout)

out1 = F.conv1d(features, weights, bias, stride=1, padding=0,
        groups=num_instances) # fails for N=0

features = features.reshape(num_instances, C, num_points)
weights = weights.reshape(num_instances, Cout, C)
bias = bias.reshape(num_instances, Cout, 1)
out2 = torch.einsum("nck,ndc->ndk", features, weights) + bias

print(out2.shape)
print(torch.abs(out1.reshape(out2.shape)-out2).max())

I remember that I had a confusion earlier when reviewing the code using group conv

[ppwwyyxx]

Should be fixed by #2998