Cannot load pretrained weights into a modified network about admd HOT 7 CLOSED

I did not check, but I guess var.assign only creates the assignment operation.
That is, you have to run that operation after you create it, like this

assign_op = var.assign(res)
sess.run(assign_op)  # also assign_op.op.run()

from admd.

That's strange because in PyTorch with the same network, optimizer, parameters and loss, 4 channels (RGB+Depth) gives better results than just 3 (RGB only) and that's what's expected, but with your code in tensorflow accuracy is lower with 4 channels...I think I will have no choice but to compare in PyTorch and extrapolate from there...

from admd.

This issue should not be closed because your code still don’t work with RGB+D concatenation...It should work like in pytorch (improve the performance and not decreasing them).

from admd.

Hi. This code refers to our paper. We never ever perform rgb+d concatenation there. I have been happy to help you with your code, but I cannot be in charge of solving issues with your modified version of my code.

from admd.

Ok but do you have any idea how its possible that your code gives lower performances with RGB+Depth concatenate than just RGB and Depth while in PyTorch its better ?

from admd.

There can be many reasons I think

1. are you sure you are assigning the rgb weights correctly?
2. random uniform initialization may not be a good choice, you can leave the default init which is automatically run in the solver (here)
3. make sure you are training the new 4D filters and that they are in the list of trainable variable
   3a) make sure the loss is optimized wrt such variables (here)
4. what kind of input are you using? depth in meters? here I am using HHA encoding, which is 3 channels.
   4a) if you normalize depth in meters in the range of RGB this is probably not going to work, you are squeezeing too much. Try to visualized the nromalized data.

Hope this is useful.

from admd.

Thanks a lot I’m gonna check all those possible issues, when the problem will be solved I will write de solution here!

from admd.