Automatically converting to float 64?

vision
1

I am working with object detrection, when I declare the dataset class, I make the boxes to be flaot32 but when I output them using dataloaders, it becomes float64.
Also there is an error saying
Variable._ execution_engine.run_backward( 148 tensors, grad tensors _, retain_graph, create_graph, inputs, → 149 allow_unreachable=True, accumulate_grad=True) # allow_unreachable flag 150 151 RuntimeError: Found dtype Double but expected Float
I think its because of this
Dataset class(taken from 🐠 Reef- Starter Torch FasterRCNN Train [LB=0.416] | Kaggle)

class ReefDataset:

    def __init__(self, df, transforms=None):
        self.df = df
        self.transforms = transforms

    def can_augment(self, boxes):
        """ Check if bounding boxes are OK to augment
        
        
        For example: image_id 1-490 has a bounding box that is partially outside of the image
        It breaks albumentation
        Here we check the margins are within the image to make sure the augmentation can be applied
        """
        
        box_outside_image = ((boxes[:, 0] < 0).any() or (boxes[:, 1] < 0).any() 
                             or (boxes[:, 2] > 1280).any() or (boxes[:, 3] > 720).any())
        return not box_outside_image

    def get_boxes(self, row):
        """Returns the bboxes for a given row as a 3D matrix with format [x_min, y_min, x_max, y_max]"""
        
        boxes = pd.DataFrame(row['annotations'], columns=['x', 'y', 'width', 'height']).astype(float).values
        
        # Change from [x_min, y_min, w, h] to [x_min, y_min, x_max, y_max]
        boxes[:, 2] = boxes[:, 0] + boxes[:, 2]
        boxes[:, 3] = boxes[:, 1] + boxes[:, 3]
        return boxes
    
    def get_image(self, row):
        """Gets the image for a given row"""
        
        image = cv2.imread(f'{BASE_DIR}/{row["image_path"]}', cv2.IMREAD_COLOR)
        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB).astype(np.float32)
        image /= 255.0
        return image
    
    def __getitem__(self, i):

        row = self.df.iloc[i]
        image = self.get_image(row)
        boxes = self.get_boxes(row)
        
        n_boxes = boxes.shape[0]
        
        # Calculate the area
        area = (boxes[:, 3] - boxes[:, 1]) * (boxes[:, 2] - boxes[:, 0])
        
        
        target = {
            'boxes': torch.as_tensor(boxes, dtype=torch.float32),
            'area': torch.as_tensor(area, dtype=torch.float32),
            
            'image_id': torch.tensor([i]),
            
            # There is only one class
            'labels': torch.ones((n_boxes,), dtype=torch.int64),
            
            # Suppose all instances are not crowd
            'iscrowd': torch.zeros((n_boxes,), dtype=torch.int64)            
        }

        if self.transforms and self.can_augment(boxes):
            sample = {
                'image': image,
                'bboxes': target['boxes'],
                'labels': target['labels']
            }
            sample = self.transforms(**sample)
            image = sample['image']
            
            if n_boxes > 0:
                target['boxes'] = torch.stack(tuple(map(torch.tensor, zip(*sample['bboxes'])))).permute(1, 0)
        else:
            image = ToTensorV2(p=1.0)(image=image)['image']

        return image, target

    def __len__(self):
        return len(self.df)

image, targets = ds_train[2200]
targets['boxes']
//output
tensor([[510.0000, 597.7778, 532.5000, 626.6667],
        [560.0000, 173.3333, 575.6250, 201.1111],
        [386.2500, 497.7778, 410.6250, 546.6667],
        [348.1250, 403.3333, 373.7500, 447.7778],
        [610.6250, 531.1111, 629.3750, 558.8889],
        [515.0000, 568.8889, 537.5000, 608.8889],
        [300.0000, 307.7778, 339.3750, 382.2222],
        [613.1250, 430.0000, 648.1250, 474.4444],
        [290.0000, 346.6667, 314.3750, 412.2222],
        [581.2500, 488.8889, 596.2500, 515.5555],
        [693.7500, 304.4445, 712.5000, 345.5555],
        [675.6250, 262.2222, 693.7500, 300.0000],
        [378.7500, 152.2222, 398.7500, 192.2222]], dtype=torch.float64)

Help needed!

2

I don’t know which code path is taken in your __getitem__ method, but since ToTensorV2 uses torch.from_numpy directly (and numpy uses float64 by default) I would assume this code path keeps the dtype as float64.

3

I checked my images, data type its float32, but the bbox float64

4

OK, I see. In that case check the dtype before calling self.transforms on the boxes (if used) and afterwards. If this doesn’t transform them back to float64, check the target['boxes'] = torch.stack(...) line.

1 Like