the example code causes bug

[Baifeng76]

Hi, there! I was running the code provided in your README about running one sequence with full-frame image with eye_net_m, but I got an error. This output is as follows:

$ python3 main.py --sequence 1 --mode org --model eye_net_m --dataset openEDS --pytorch_model_path model_weights/eye_net_m.pkl --device cpu --preview
Infer openEDS/S_1 (1/151)
openEDS/S_1/0.png
eye segmentation baseline (400, 640)
Traceback (most recent call last):
File "main.py", line 253, in
main()
File "main.py", line 246, in main
output_video_path=video_name,
File "/edgaze/deepvog/inferer.py", line 148, in process
Y_batch = self.model.predict(framename)
File "/edgaze/edgaze/eye_segmentation.py", line 147, in predict
img = skimage.color.rgb2gray(img).squeeze().astype(numpy.uint8)
File "/miniconda3/envs/python/lib/python3.7/site-packages/skimage/_shared/utils.py", line 394, in fixed_func
return func(*args, **kwargs)
File "/miniconda3/envs/python/lib/python3.7/site-packages/skimage/color/colorconv.py", line 875, in rgb2gray
rgb = _prepare_colorarray(rgb)
File "/miniconda3/envs/python/lib/python3.7/site-packages/skimage/color/colorconv.py", line 140, in _prepare_colorarray
raise ValueError(msg)
ValueError: the input array must have size 3 along channel_axis, got (400, 640)
Exception ignored in: <function VideoManager.del at 0x7f40fd0e7680>
Traceback (most recent call last):
File "/edgaze/deepvog/visualisation.py", line 141, in del
self.vwriter.close()
File "/miniconda3/envs/python/lib/python3.7/site-packages/skvideo/io/ffmpeg.py", line 430, in close
if self._proc is None: # pragma: no cover
AttributeError: 'FFmpegWriter' object has no attribute '_proc'

but the other example runs well. Could you please fix it?

[YuFengUofR]

did you change the code or input images at all? it seems like the input image dimension of function rgb2gray is incorrect

[Baifeng76]

No, I didn't change the code or the input images. Now I find that predict_filtering of EyeSegmentation expands the dims of input img when the dim of it equals to 2, and seems not to use rgb2gray. But predict function doesn't do that. Maybe this caused the bug.

[YuFengUofR]

that might be the case, try to make data dimensions consistent.

[Baifeng76]

Thanks, now the demo code runs well.
I want to test the accuracy and speed of your model. Could you provide your testing code? Thanks!

[Ahmet-Kaplan]

Hi @Baifeng76 ,

I got same error as yours. Can you share the working code ?
Thanks,

[Baifeng76]

Hi @Baifeng76 ,

I got same error as yours. Can you share the working code ? Thanks,

Hi @Ahmet-Kaplan ,

I changed the line 147 of eye_segmentation.py into:

    if img.dim == 2:
        img = img.astype(numpy.unit8)
    else:
        img = skimage.color.rgb2gray(img).squeeze().astype(numpy.uint8)

and it works. Hope this can help you.

[Ahmet-Kaplan]

Yes, with small typo corrections, it works.
if img.ndim == 2:
img = img.astype (numpy.uint8)
else:
img = skimage.color.rgb2gray (img).squeeze ().astype (numpy.uint8)