Error while using transforms ToPILImage and RandomResizedCrop

vision
1

Hi,

I’m trying to use the transforms ToPILImage and RandomResizedCrop and run into the following errors.

For ToPILImage:


  File "/home/usr/anaconda3/lib/python3.6/site-packages/torchvision-0.2.0-py3.6.egg/torchvision/transforms/functional.py", line 139, in to_pil_image
    raise TypeError('Input type {} is not supported'.format(npimg.dtype))

TypeError: Input type float32 is not supported

For RandomResizedCrop:


  File "/home/usr/anaconda3/lib/python3.6/site-packages/torchvision-0.2.0-py3.6.egg/torchvision/transforms/transforms.py", line 396, in get_params
    area = img.size[0] * img.size[1]

TypeError: 'int' object is not subscriptable

The code to reproduce this error for ToPILImage:

import numpy as np
import torch
from torchvision import transforms
from torch.utils.data import DataLoader, Dataset

class given_data(Dataset):

    def __init__(self,images,labels, transform = None):
        self.images = images
        self.labels = labels
        self.transform = transform
    
    def __len__(self):
        return len(self.labels)
    
    def __getitem__(self,indx):
        imgs = self.images[indx]
        labls = self.labels[indx]
        
        if self.transform:
            imgs = self.transform(np.float32(imgs))
        return (imgs,labls)
    
X_inp = np.random.rand(100,75,75,2)
Y_inp = np.random.rand(100)


txfm = transforms.Compose([transforms.ToPILImage()])

transformed_dataset = given_data(X_inp,Y_inp,transform= txfm)

print(transformed_dataset[2])

For RandomResizedCrop simply replace the following line:

txfm = transforms.Compose([transforms.RandomResizedCrop((80,80),ratio=(1,1))])

I’ve installed torchvision from source.

Can someone please help me out?

2

Hi @barrel-roll,

The reason why you got TypeError: Input type float32 is not supported is because the default transforms.ToPILImage() has mode=None and the third dimension in your synthetic data is 2 (which is usually set to 3 for color images), leading to the TypeError here https://github.com/pytorch/vision/blob/master/torchvision/transforms/functional.py#L138. (you can follow the source code from here https://github.com/pytorch/vision/blob/master/torchvision/transforms/functional.py#L81 where the pic you gave is (75, 75, 2) and mode is None).

One solution is to set the data shape to (75, 75, 3) and change imgs = self.transform(np.float32(imgs)) to imgs = self.transform(np.uint8(imgs)).

What the ToPILImage() does is basically to call Image.fromarray(), so you might want to have a look at https://pillow.readthedocs.io/en/5.0.0/reference/Image.html.

3 Likes
3

Hi @KaiyangZhou ,

Thanks, that does explain why it ToPILImage didn’t work. Unfortunately I’m dealing with Radar data for the Iceberg challenge hence, so neither can I convert it to a 3 channel image nor have uint8 format.

Did you find why RandomResizedCrop throws an error?

4

You can directly transform Radar data into torch tensor, but you need to perform operations (e.g. resize) beforehand (i.e. implement your operations manually).

I could not reproduce the error with RandomResizedCrop.

5

Thanks. I’ll try the manual implementation then.

Regarding RandomResizedCrop, what’s your version of torchvision? Did you build from source? I rechecked and it’s still giving me error.

6

My version is 0.2.0.

Try transforms.RandomResizedCrop(80, ratio=(1,1))?

7

My version is the same. And nope, still the same error.

This is really weird…

8

I also faced this problem, in my case I read image from cv2, then I used transforms.RandomResziedCro()p. I solved this by adding transforms.ToPILImage() before transforms.RandomResziedCrop().