CocoDetection with transforms not working

vision
1

My goal is to train a pre-trained object segmentation model using my own dataset with its own classes. So, I created my own dataset using the COCO Dataset format.

Here below, you can see that I am trying to create a Dataset using the function CocoDetection.

import torchvision.datasets as dset

def get_transform():
    custom_transforms = []
    custom_transforms.append(T.ToTensor())
    return T.Compose(custom_transforms)

coco_test = dset.CocoDetection(root = root_dir,
                                annFile = test_json,
                                transforms=get_transform())

# Test if dataset is working
for i in range(len(coco_train)):
  sample = coco_train[i]
  print(sample[0])

  if i == 3:
    break

At first, I tried without applying any transformations at all and it worked. However, after I added a transform on the function, it did not work. I checked the error and it says that the call takes 2 positional argument, but 3 were given. But, I inspected the source code, and I am giving it 2 arguments. I don’t know what’s the problem. I don’t know what to do next. Please help. Here’s the full error. By the way, I am using Google Colab.

---------------------------------------------------------------------------

TypeError                                 Traceback (most recent call last)

<ipython-input-13-bc0dddb4b6c5> in <module>()
      1 for i in range(len(coco_train)):
----> 2   sample = coco_train[i]
      3   print(sample[0])
      4 
      5   if i == 3:

/usr/local/lib/python3.6/dist-packages/torchvision/datasets/coco.py in __getitem__(self, index)
    116         img = Image.open(os.path.join(self.root, path)).convert('RGB')
    117         if self.transforms is not None:
--> 118             img, target = self.transforms(img, target)
    119 
    120         return img, target

TypeError: __call__() takes 2 positional arguments but 3 were given
2

From the docs for CocoDetection:

  • transform ( callable , optional ) – A function/transform that takes in an PIL image and returns a transformed version. E.g, transforms.ToTensor
  • target_transform ( callable , optional ) – A function/transform that takes in the target and transforms it.
  • transforms ( callable , optional ) – A function/transform that takes input sample and its target as entry and returns a transformed version.

You could either pass the transform and target_transform separately or use a transforms function, which accepts the image and target.

3

I don’t get it. So, is there something wrong to what I did? I read other codes and they also created a list for transforms too. And why would the number of positional arguments I passed be 3, when it is clearly two? Do you think maybe it is just not maintained to the latest version?

4

In other code snippets the transformation was most likely only applied to the data, so a single transform argument was passed.
For the CocoDetection dataset, you are transforming the data and target, so you should either pass the transformations separately or provide a transformation function, which accepts the data and the target as arguments.

The self argument is also counted in Python.

No, I don’t think so.

EDIT: Note that these arguments are optional, so you can also pass your transformation as transform=get_transform() only to skip the target transformation.

1 Like
5

Ohh right, the self is also counted as an argument. Thanks for the reminder!
Anyway, I just solved my problem. It appears that I am using the wrong CocoDetection Dataset class. I orginally used:

import torchvision.datasets as dset
# and then
coco_test = dset.CocoDetection(root = root_dir,
                                annFile = test_json,
                                transforms=get_transform())

When what I should’ve used is:

from torchvision.datasets.coco import CocoDetection
# and then
coco_test = CocoDetection(root = root_dir,
                                annFile = test_json,
                                transforms=get_transform())

And it worked!

So, it appears that there are two CocoDetection classes. One in torchvision.datasets, another in torchvision.datasets.coco. Should I inform them about this duplicate code? And if so, how should I inform them about this?

6

Both classes are pointing to the same implementation and I’m wondering how you solved the issue, but it’s good to hear it’s working. :slight_smile:

7

I think it may have something to do with the first implementation inheriting the class VisionDataset. While on the second one, the implementation inherits the data.Datasets.