Problems with lmdb when there're multiple inputs of image and corresponding labels of different size

[Ben]

For example, I have inputs data1, data2, and corresponding labels label1 and label2.  label are pixelwise labeling of images.

And I met a problem:

case 1 multiple input of different size:

       data1 + label1 ----> loss1  (* 1.)

       data2 + label2 ----> loss2  (* 0.)

case 2 multiple input of same  size:

       data1+ label1 ---->  loss1  (* 1.)

       data1+ label1-----> loss1   (* 0.)

case 3 single input:

       data1+ label1 ------> loss1

 

In my test experiment, the effect of case 2 and case 3 is same.

But I found the loss of case 1 is quite strange, So I try to test the net. In case 1, I set loss_weight  0. for loss2, and observe the change of loss1 compared to case 3.  And the losses are different at the 6th image. I try to monitor the blob mean of convolution layers and loss layers.
In forward passing, all the convolution blobs has same mean.

In the loss layer, I try to get the mean of bottom[1], i.e. the label blob. and find that

mean of "label1" in case 1 is different from the mean of "label1" in case 3?

I think this is why the loss is different. what's happening, is it a bug ? Or I misused something.

 

[Ben]

It's weird that sometimes the mean of label blob is right, sometimes it's wrong. What's going on ?