Zero-sized dimensions are not supported in PyTorch about ganimation HOT 7 CLOSED

had same issue, it turns out it was the problem with images with multiple faces. It was fixed in prepare_au file with:

    for filepath in tqdm(filepaths):
        content = np.loadtxt(filepath, delimiter=', ', skiprows=1)
        if content.ndim == 1:
            data[os.path.basename(filepath[:-4])] = content[2:19]
        if content.ndim == 2:
            data[os.path.basename(filepath[:-4])] = content[0][2:19]
    return data

from ganimation.

Seems there is something wrong with your dataset. Check that the dimensions of each element of a sample are always the same.

from ganimation.

when the test sample size is 1, python will raise this error, you can add a special treat where there is only one test or train samples

from ganimation.

@xiaoiker check if there are images of zero size

from ganimation.

I checked our my images with the following code, found no zero size image. There are around 200K imgs.

input_path ='/homeappl/home/imgs'
files = glob.glob(input_path+'*.j*')
for n, img in enumerate(files):# /*.*
	try:
		var_img = cv2.imread(img) 
		height, weight, C = var_img.shape
		if (height != 128)|(weight != 128)|(C != 3):
			os.remove(img)
			print('Deleting file: {}'.format(img))
		else if var_img is None:
			os.remove(img)
			print('Deleting file: {}'.format(img))				

		print(n)
	
	except:
		os.remove(img)
		print('Deleting file: {}'.format(img))

I also checked the AUs files, found no abnormal. But still not work.

from ganimation.

@xiaoiker when load data in dataset_au.py, real_cond or desired_cond may have zero-sized dimensions, you can add a special treat, and I meet this problem and now success to train.

from ganimation.

@xiaoiker I have try the way proposed by (AidasK commented on 4 Sep),and This method is also applicable to the Celeba Dataset.

from ganimation.