fix mrcnn_bbox_loss

[keineahnung2345]

According to #786 and the paper of Fast R-CNN, L_loc should be the sum of smooth_l1_loss of its four items.

[waleedka]

@keineahnung2345 I couldn't find that reference in the paper! Can you point me to the specific section? Also, please see my comments on #786 . I think we do need to update the SmoothL1 loss, but probably not in this particular way.

[keineahnung2345]

I am referring to the following formula in Fast R-CNN:

just as @jnd77 says.
I've checked your comment and revise the denominator to N_reg.

[jnd77]

@keineahnung2345 thanks for looking at the issue.
However, I have the feeling we are mixing the 2 localisation losses:

The Fast-RCNN one (mrcnn_bbox_loss) which you first commit did correct.

The Faster-RCNN one (rpn_bbox_loss) which @waleedka mentioned in #786 (where the denominator is Nreg).

So I believe the first commit was the correct one.

[keineahnung2345]

@jnd77 You are right. I will revert to the first version after @waleedka confirm it.
In the paper of Mask R-CNN, it says:

Formally, during training, we define a multi-task loss on
each sampled RoI as L = Lcls + Lbox + Lmask. The classification loss Lcls and bounding-box loss Lbox are identical as those defined in [12].

In which [12] links to Fast R-CNN.
And reference to benchmark - fast r-cnn box loss, the denominator should be the count of proposals.

[jimmy15923]

Hi @keineahnung2345 !
Thanks for updating the bbox loss, I am wondering if this change affects the training?

[yhcccc]

@keineahnung2345 , I modified it as you proposed, but got error:

Traceback (most recent call last):
  File "samples/coco/coco.py", line 464, in <module>
    model_dir=args.logs)
  File "/data1/gaoxi/yh/Tensorflow/lib/python3.6/site-packages/mask_rcnn-2.1-py3.6.egg/mrcnn/model.py", line 1955, in __init__
  File "/data1/gaoxi/yh/Tensorflow/lib/python3.6/site-packages/mask_rcnn-2.1-py3.6.egg/mrcnn/model.py", line 2138, in build
  File "/data1/gaoxi/yh/Tensorflow/lib/python3.6/site-packages/keras/engine/base_layer.py", line 457, in __call__
    output = self.call(inputs, **kwargs)
  File "/data1/gaoxi/yh/Tensorflow/lib/python3.6/site-packages/keras/layers/core.py", line 687, in call
    return self.function(inputs, **arguments)
  File "/data1/gaoxi/yh/Tensorflow/lib/python3.6/site-packages/mask_rcnn-2.1-py3.6.egg/mrcnn/model.py", line 2137, in <lambda>
  File "/data1/gaoxi/yh/Tensorflow/lib/python3.6/site-packages/mask_rcnn-2.1-py3.6.egg/mrcnn/model.py", line 1212, in mrcnn_bbox_loss_graph
  File "/data1/gaoxi/yh/Tensorflow/lib/python3.6/site-packages/tensorflow/python/ops/math_ops.py", line 856, in binary_op_wrapper
    y = ops.convert_to_tensor(y, dtype=x.dtype.base_dtype, name="y")
  File "/data1/gaoxi/yh/Tensorflow/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 611, in convert_to_tensor
    as_ref=False)
  File "/data1/gaoxi/yh/Tensorflow/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 676, in internal_convert_to_tensor
    ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)
  File "/data1/gaoxi/yh/Tensorflow/lib/python3.6/site-packages/tensorflow/python/framework/constant_op.py", line 121, in _constant_tensor_conversion_function
    return constant(v, dtype=dtype, name=name)
  File "/data1/gaoxi/yh/Tensorflow/lib/python3.6/site-packages/tensorflow/python/framework/constant_op.py", line 102, in constant
    tensor_util.make_tensor_proto(value, dtype=dtype, shape=shape, verify_shape=verify_shape))
  File "/data1/gaoxi/yh/Tensorflow/lib/python3.6/site-packages/tensorflow/python/framework/tensor_util.py", line 364, in make_tensor_proto
    raise ValueError("None values not supported.")
ValueError: None values not supported.