This was just testing the get_params method.
Now add the actual cropping to your code:
crop = MyRandomCrop((256, 256))
for i in range(10):
img = transforms.ToPILImage()(torch.randn(3, 600, 600))
print(crop(img).size)
this is the output
Using 155 153 256 256
(256, 256)
Using 155 153 256 256
(256, 256)
Using 78 168 256 256
(256, 256)
Using 78 168 256 256
(256, 256)
Using 234 6 256 256
(256, 256)
Using 234 6 256 256
(256, 256)
Using 39 179 256 256
(256, 256)
Using 39 179 256 256
(256, 256)
Using 214 317 256 256
(256, 256)
Using 214 317 256 256
(256, 256)
so yes it is working here
Still strange that it’s not working in your code.
I’ve tried your get_data_loader_folder method and it’s also working.
Could you try to set the crop size to something really small, e.g. 10?
I printed the (self.counter, resample) inside the function … look what it printed
(1, True)
Using 0 0 256 256
(2, False)
Using 0 0 256 256
(3, True)
Using 0 0 256 256
(4, False)
Using 0 0 256 256
(1, True)
Using 0 0 256 256
(2, False)
Using 0 0 256 256
(3, True)
Using 0 0 256 256
(4, False)
Using 0 0 256 256
(1, True)
Using 0 0 256 256
(2, False)
Using 0 0 256 256
(3, True)
Using 0 0 256 256
(4, False)
Using 0 0 256 256
(1, True)
Using 0 0 256 256
(2, False)
Using 0 0 256 256
(3, True)
Using 0 0 256 256
(4, False)
Using 0 0 256 256
(5, True)
Using 0 0 256 256
(5, True)
(5, True)
Using 0 0 256 256
Using 0 0 256 256
(5, True)
Using 0 0 256 256
(5, True)
Using 0 0 256 256
(5, True)
Using 0 0 256 256
(5, True)
(5, True)
Using 0 0 256 256
Using 0 0 256 256
(6, False)
Using 0 0 256 256
(6, False)
Using 0 0 256 256
(6, False)
Using 0 0 256 256
(6, False)
Using 0 0 256 256
(6, False)
Using 0 0 256 256
(6, False)
Using 0 0 256 256
(6, False)
Using 0 0 256 256
(6, False)
Using 0 0 256 256
(7, True)
Using 0 0 256 256
(5, True)
Using 0 0 256 256
(5, True)
Using 0 0 256 256
(5, True)
Using 0 0 256 256
(5, True)
Using 0 0 256 256
(5, True)
Using 0 0 256 256
(5, True)
Using 0 0 256 256
(5, True)
Using 0 0 256 256
(5, True)
Using 0 0 256 256
(6, False)
Using 0 0 256 256
(6, False)
(6, False)
Using 0 0 256 256
Using 0 0 256 256
(6, False)
Using 0 0 256 256
(6, False)
Using 0 0 256 256
(6, False)
Using 0 0 256 256
(6, False)
Using 0 0 256 256
(6, False)
Using 0 0 256 256
(7, True)
Using 0 0 256 256
here it is
Using 221 55 10 10
Using 221 55 10 10
Using 243 87 10 10
Using 243 87 10 10
Using 89 94 10 10
Using 89 94 10 10
Using 145 83 10 10
Using 145 83 10 10
Using 195 113 10 10
Using 195 113 10 10
Using 78 148 10 10
Using 78 148 10 10
Using 171 70 10 10
Using 171 70 10 10
Using 125 57 10 10
Using 125 57 10 10
Using 111 225 10 10
Using 79 10 10 10
Using 206 55 10 10
Using 12 192 10 10
Using 158 132 10 10
Using 143 62 10 10
Using 23 71 10 10
Using 163 115 10 10
Using 111 225 10 10
Using 79 10 10 10
Using 12 192 10 10
Using 206 55 10 10
Using 158 132 10 10
Using 143 62 10 10
Using 163 115 10 10
Using 23 71 10 10
Using 174 39 10 10
ok now it worked
but there are two things weird
why it worked with a small size and not 256
also, if you check the print outs, you can see that they are consistent and then something weird happens
I tried with 255
look what happened
Using 1 0 255 255
Using 1 0 255 255
Using 1 0 255 255
Using 1 0 255 255
Using 0 0 255 255
Using 0 0 255 255
Using 0 1 255 255
Using 0 1 255 255
Using 1 1 255 255
Using 1 1 255 255
Using 1 1 255 255
Using 1 1 255 255
Using 0 1 255 255
Using 0 1 255 255
Using 0 0 255 255
Using 0 0 255 255
Using 0 0 255 255
Using 0 0 255 255
Using 0 1 255 255
Using 0 1 255 255
Using 1 0 255 255
Using 1 1 255 255
Using 1 1 255 255
Using 1 1 255 255
Using 0 0 255 255
Using 0 0 255 255
Using 0 1 255 255
Using 1 1 255 255
Using 0 1 255 255
Using 1 0 255 255
Using 1 1 255 255
Using 1 1 255 255
Using 0 0 255 255
Using 1 0 255 255
Using 0 0 255 255
Using 0 0 255 255
Using 0 0 255 255
Using 0 0 255 255
Using 0 1 255 255
Using 1 1 255 255
Using 0 0 255 255
Using 0 1 255 255
Using 0 0 255 255
Using 0 0 255 255
Using 0 0 255 255
Using 1 1 255 255
Using 1 0 255 255
Using 0 0 255 255
Using 0 0 255 255
Using 1 0 255 255
as far as I understand and notice, I believe that the resizing happens before the cropping
Exactly! Was going to post the same.
You are resizing the images to [256, 256] so a crop won’t do anything afterwards.
Change your resize shape or crop size.
yes exactly, as i assumed
I removed the resize and it worked
Using 28 48 256 256
Using 28 48 256 256
Using 0 127 256 256
Using 0 127 256 256
Using 115 330 256 256
Using 115 330 256 256
Using 224 279 256 256
Using 224 279 256 256
Using 41 62 256 256
Using 41 62 256 256
Using 18 126 256 256
Using 18 126 256 256
Using 180 271 256 256
Using 180 271 256 256
Using 217 372 256 256
Using 217 372 256 256
Using 12 120 256 256
Using 15 58 256 256
Using 36 4 256 256
Using 40 44 256 256
Using 23 25 256 256
Using 7 143 256 256
Using 4 85 256 256
Using 18 57 256 256
Using 36 4 256 256
Using 40 44 256 256
Using 15 58 256 256
Using 18 57 256 256
Using 7 143 256 256
Using 12 120 256 256
Using 4 85 256 256
Using 23 25 256 256
Using 42 11 256 256
Using 134 16 256 256
Using 7 283 256 256
Using 41 335 256 256
Using 220 381 256 256
Using 129 228 256 256
Using 37 206 256 256
Using 60 177 256 256
Using 340 358 256 256
Using 7 283 256 256
Using 220 381 256 256
Using 41 335 256 256
Using 134 16 256 256
Using 129 228 256 256
Using 60 177 256 256
Using 340 358 256 256
Using 37 206 256 256
but again here lies the problem, that the indices become inconsistent at some point
This might be due to using multiple workers.
and does this have a solution by any means ?
Let me think about a solution and make sure I understand the issue properly.
You want a batch of tensors where two neighboring samples were cropped using the same indices.
Also your dataset might be shuffled, so that the order of the images does not matter.
Multiple workers should also be used.
Is that correct?
concerning the shuffling where I already asked here, I solved it by doing like this
def shuffle_pairs(images):
pairs = [[images[z+y] for y in range(2)] for z in range(len(images)-1)]
random.shuffle(pairs)
shuffled = []
for x in pairs:
for y in x:
shuffled.append(y)
return shuffled
class ImageFolder(data.Dataset):
def __init__(self, root, transform=None, return_paths=False,
loader=default_loader):
#imgs = sorted(make_dataset(root))
# shuffle implicit pairs
if "test" in root:
imgs = sorted(make_dataset(root))
else:
imgs = shuffle_pairs(sorted(make_dataset(root)))
if len(imgs) == 0:
raise(RuntimeError("Found 0 images in: " + root + "\n"
"Supported image extensions are: " +
",".join(IMG_EXTENSIONS)))
self.root = root
self.imgs = imgs
self.transform = transform
self.return_paths = return_paths
self.loader = loader
def __getitem__(self, index):
path = self.imgs[index]
img = self.loader(path)
if self.transform is not None:
img.save('x_a1.png')
img = self.transform(img)
vutils.save_image(img, 'x_a2.png', nrow=1)
#exit()
if self.return_paths:
return img, path
else:
return img
def __len__(self):
return len(self.imgs)
so i didn’t use any sampler or something related to a dataloader … I just shuffled the list after reading it from the folder
so this works fine for me now
what i want, is that for every 2 samples, they have the same cropping
so sample 1 and 2, will be cropped from the same position
sample 3 and 4
sample 5 and 6
and so on
Was this solution working properly?
I have the feeling that the mixup of the crop sizes might be due to some racing conditions of the workers, which would also influence the order of your samples.
without trying the cropping, i think i have no problem with this solution, it is working okay
Could you check if with the indices please?
I.e. add your load logic to this crop approach and print or store the indices you are expecting.
I don’t see a way of getting the indices here before and after shuffling
import torch
import torchvision.transforms.functional as TF
from torchvision import transforms
import random
class MyRandomCrop(transforms.RandomCrop):
def __init__(self, size, padding=0, pad_if_needed=False):
super(MyRandomCrop, self).__init__(size, padding, pad_if_needed)
self.counter = 0
self.crop_indices = []
def __call__(self, img):
if self.padding > 0:
img = TF.pad(img, self.padding)
# pad the width if needed
if self.pad_if_needed and img.size[0] < self.size[1]:
img = TF.pad(img, (int((1 + self.size[1] - img.size[0]) / 2), 0))
# pad the height if needed
if self.pad_if_needed and img.size[1] < self.size[0]:
img = TF.pad(img, (0, int((1 + self.size[0] - img.size[1]) / 2)))
resample = self.counter % 2 == 0
self.counter += 1
#print(self.counter, resample)
if resample:
self.crop_indices = self.get_params(img, self.size)
i, j, h, w = self.crop_indices
print('Using {} {} {} {}'.format(i, j, h, w))
return TF.crop(img, i, j, h, w)
def shuffle_pairs(images):
pairs = [[images[z+y] for y in range(2)] for z in range(len(images)-1)]
random.shuffle(pairs)
shuffled = []
for x in pairs:
for y in x:
shuffled.append(y)
return shuffled
images = []
crop = MyRandomCrop((256, 256))
for i in range(10):
img = transforms.ToPILImage()(torch.randn(3, 600, 600))
images.append(img)
shuffled_images = shuffle_pairs(images)
for i in (shuffled_images):
print(crop(i).size)