The error Unsupported dtype object

[yuzhesu]

when I run the generate_result.py ,and set up correct, it appear The error Unsupported dtype object! How can I solve this problem?

[SeitaroShinagawa]

Thanks for your report.
Could you check batch[i][1] in line 111 by "print(type(batch[i][1]))"?
It should be a numpy array, but it seems None in your case.
If it is None, check whether the dataset path and mask dataset path is correct or not.

[yuzhesu]

I have already check batch[i][1] in line 111 by "print(type(batch[i][1])), it is None.
But I set val_dataset and mask_path correct.
Or #test_dataset I should also set it?

[yuzhesu]

I have already change test_place2 in Edit common/paths.py
But when I run the generate_result.py , I still can't get the test pictures, If I want to get the test pictures, Where should I set the path?

[SeitaroShinagawa]

You don't need use test_dataset.
This is a simple example to load dataset. Could you try to execute the following code?

#!/usr/bin/env python

import datasets
from common import paths

val_dataset = getattr(datasets, "place2_test")(paths.val_place2, mask_path="mask/256", resize_to=256, crop_to=256)
val_iter = chainer.iterators.SerialIterator(val_dataset, 1)
batch = val_iter.next()

print(type(batch[0][1]))
#correct return --> <class 'numpy.ndarray'>

[yuzhesu]

When I execute the following code, it show <class 'NoneType'>
How can I solve this problem？

[SeitaroShinagawa]

There are 2 possibilities, paths.val_place2 is wrong or mask_path is wrong.
Judging from your case, loading mask has failed.
Could you show the result of "ls" and "ls mask" command in the root directory of this repository?

[yuzhesu]

I think paths.val_place2 is setting correct.
Where can mask_path be set？
Sorry, what is "ls" and "ls mask" meaning for?

[SeitaroShinagawa]

Sorry, I want to see where is your mask path.
If your mask path is correct, you can do the following commands:

>>> import glob
>>> maskkeys = glob.glob("mask/256/*.bmp")
>>> print(maskkeys)
['mask/256/mask_256_00.bmp', 'mask/256/mask_256_07.bmp', 'mask/256/mask_256_09.bmp', 'mask/256/mask_256_08.bmp', 'mask/256/mask_256_06.bmp', 'mask/256/mask_256_01.bmp', 'mask/256/mask_256_02.bmp', 'mask/256/mask_256_05.bmp', 'mask/256/mask_256_04.bmp', 'mask/256/mask_256_03.bmp']

Could you try this?

[yuzhesu]

Sorry,Is this result wrong?

[SeitaroShinagawa]

Oh, I think glob has some problem with windows.
Can you load the mask from the paths? For example,

from PIL import Image
img = Image.open(maskkeys[0])

[yuzhesu]

sorry , it has some problems.

[SeitaroShinagawa]

"mask/256*.bmp" is not a list but string. I mean you should try

>>> import glob
>>> from PIL import Image
>>> maskkeys = glob.glob("mask/256/*.bmp")
>>> img = Image.open(maskkeys[0])

Does it work?

[yuzhesu]

Yes, it can work.

So how do I generate the result correctly？
Thanks

[SeitaroShinagawa]

I suppose cv2.imread has something wrong in your environment.
https://github.com/SeitaroShinagawa/chainer-partial_convolution_image_inpainting/blob/master/datasets/place2.py#L109

Could you try to replace this line as follows?

#mask = cv2.imread(idM, cv2.IMREAD_GRAYSCALE)
mask = np.array(Image.open(idM)).astype("f")

However, I wonder why you can excecute train.py if cv2.imread is the cause...

[yuzhesu]

I have already replace this line as follows, but it still have error

Because I didn't perform train.py , I directly download model500000.npz to execute generate_result.py.

[SeitaroShinagawa]

The error occurred in loading mask from your mask path "idM."
Surprisingly, idM has become changed to 'C:\Users\La' in your case.
Could you try "print(idM)" before the line 110 in place2.py?

[yuzhesu]

I have already "print(idM)" before the line 110 in place2.py, Show the same result 'C:\Users\La'.

I think error is on place2.py line103 [:len(self.trainAkey)] here,
If I change line 103 to idM = self.maskkey[np.random.randint(0,len(self.maskkey))]
it will show

How can I correct [:len(self.trainAkey)] ?

[SeitaroShinagawa]

Umm, can you do "print(self.maskkey)" before the line 110? It seems the cause you have.

[yuzhesu]

Here!

[SeitaroShinagawa]

Oh, you are right. The error should be [:len(self.trainAkey)]. Could you remove it?

[SeitaroShinagawa]

I have fixed it (remove [:len(self.trainAkey)]) and it works.

ad4048b#diff-ab9bbf0e02017a0e3d0a81a64530a22eR98

Thank you for your patience!

[yuzhesu]

Yes, I have already remove it, the result is

If I remove [:len(self.trainAkey)] results will have an impact?
Thank you for your answer patiently.

[yuzhesu]

Sorry I have another question, If I want to get the Input result like this

How can I set?

[SeitaroShinagawa]

You want to get 1 image, is it correct?
Change the input of "batch_postprocess_images"
https://github.com/SeitaroShinagawa/chainer-partial_convolution_image_inpainting/blob/master/generate_result.py#L92
Best,

[yuzhesu]

No, I want to get the image like the paper's input image.

[SeitaroShinagawa]

My implementation does not support it. However, you can do it by replacing the mask part value with 1.0 on each channel.

You can do like here,
https://github.com/SeitaroShinagawa/chainer-partial_convolution_image_inpainting/blob/master/generate_result.py#L88

#easiest way
output_image = F.where(M_b, I_gt, Variable(xp.ones((batchsize, 3, image_size, image_size)).astype("f"))

[yuzhesu]

I understand!Thank you for your answer patiently.