Mask_rcnn: inspect_model :ValueError: Floating point image RGB values must be in the 0..1 range.

[PaulChongPeng]

run code in inspect_model.ipynb:

_ = plt.imshow(activations["input_image"][0])

then:

~/Tools/tensorflow_python3/lib/python3.5/site-packages/matplotlib/cm.py in to_rgba(self, x, alpha, bytes, norm)
    255                 if xx.dtype.kind == 'f':
    256                     if norm and xx.max() > 1 or xx.min() < 0:
--> 257                         raise ValueError("Floating point image RGB values "
    258                                          "must be in the 0..1 range.")
    259                     if bytes:
ValueError: Floating point image RGB values must be in the 0..1 range.

I check the code and find that :

input_image              shape: (1, 1024, 1024, 3)    min: -123.70000  max:  148.10001

So I modify the code :

input_image_a = activations["input_image"][0]+config.MEAN_PIXEL
input_image_a = A.astype(int)
input_image_a = A/255
log("input_image_a:",input_image_a)
_ = plt.imshow(input_image_a)

And then the image showed but there are some difference between your image in inspect_model.ipynb and mine.
My image is just the origin input image after resized and padded.
I think it should not be what you want to show.

[waleedka]

Which version of matplotlib are you using? Mine is 2.0.2 and it doesn't complain about requiring float values to be 0..1.

matplotlib.__version__

[waleedka]

And, you can use unmold_image() instead. It converts the image back from the form that the network expects to the original form (i.e. reverses the effect of mold_image())

[PaulChongPeng]

Thanks!
My matplotlib version is 2.1.0.
I uninstall it and install matplotlib version 2.0.2, and the bug is fixed.
The unmold_image() function is useful.
Could I create a pull request which modify

activations["input_image"][0]

to

modellib.unmold_image(activations["input_image"][0],config)

?

[waleedka]

Yes, please. Thank you.