Random crop error: ZeroDivisionError: float division by zero

❓ Questions and Help

I’m implementing random cropping method and my functions is the following.

class RandomCrop(object):
    def __init__(self, size):
        self.size = size

    @staticmethod
    def get_params(img, output_size):
        """Get parameters for ``crop`` for a random crop.
        Args:
            img (PIL Image): Image to be cropped.
            output_size (tuple): Expected output size of the crop.
        Returns:
            tuple: params (i, j, h, w) to be passed to ``crop`` for random crop.
        """
        w, h = img.size
        th, tw = output_size
        if w == tw and h == th:
            return 0, 0, h, w

        i = random.randint(0, h - th) #height
        j = random.randint(0, w - tw) #width
        return i, j, th, tw

    def __call__(self, image, target):
        i, j, h, w = self.get_params(image, self.size)
        image = F.crop(image, i, j ,h, w)
        target = target.crop([j, i, j+w, i+h])
        return image, target

However, I’m getting the error message as follows:

2019-04-26 15:27:16,323 maskrcnn_benchmark.trainer INFO: Start training
Traceback (most recent call last):
  File "tools/train_net.py", line 186, in <module>
    main()
  File "tools/train_net.py", line 179, in main
    model = train(cfg, args.local_rank, args.distributed)
  File "tools/train_net.py", line 85, in train
    arguments,
  File "/home/ubuntu/github/maskrcnn-benchmark/maskrcnn_benchmark/engine/trainer.py", line 67, in do_train
    loss_dict = model(images, targets)
  File "/home/ubuntu/anaconda3/envs/pytorch_p36/lib/python3.6/site-packages/torch/nn/modules/module.py", line 491, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/ubuntu/anaconda3/envs/pytorch_p36/lib/python3.6/site-packages/apex-0.1-py3.6-linux-x86_64.egg/apex/amp/_initialize.py", line 194, in new_fwd
    **applier(kwargs, input_caster))
  File "/home/ubuntu/github/maskrcnn-benchmark/maskrcnn_benchmark/modeling/detector/generalized_rcnn.py", line 52, in forward
    x, result, detector_losses = self.roi_heads(features, proposals, targets)
  File "/home/ubuntu/anaconda3/envs/pytorch_p36/lib/python3.6/site-packages/torch/nn/modules/module.py", line 491, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/ubuntu/github/maskrcnn-benchmark/maskrcnn_benchmark/modeling/roi_heads/roi_heads.py", line 39, in forward
    x, detections, loss_mask = self.mask(mask_features, detections, targets)
  File "/home/ubuntu/anaconda3/envs/pytorch_p36/lib/python3.6/site-packages/torch/nn/modules/module.py", line 491, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/ubuntu/github/maskrcnn-benchmark/maskrcnn_benchmark/modeling/roi_heads/mask_head/mask_head.py", line 77, in forward
    loss_mask = self.loss_evaluator(proposals, mask_logits, targets)
  File "/home/ubuntu/github/maskrcnn-benchmark/maskrcnn_benchmark/modeling/roi_heads/mask_head/loss.py", line 112, in __call__
    labels, mask_targets = self.prepare_targets(proposals, targets)
  File "/home/ubuntu/github/maskrcnn-benchmark/maskrcnn_benchmark/modeling/roi_heads/mask_head/loss.py", line 94, in prepare_targets
    segmentation_masks, positive_proposals, self.discretization_size
  File "/home/ubuntu/github/maskrcnn-benchmark/maskrcnn_benchmark/modeling/roi_heads/mask_head/loss.py", line 37, in project_masks_on_boxes
    scaled_mask = cropped_mask.resize((M, M))
  File "/home/ubuntu/github/maskrcnn-benchmark/maskrcnn_benchmark/structures/segmentation_mask.py", line 486, in resize
    resized_instances = self.instances.resize(size)
  File "/home/ubuntu/github/maskrcnn-benchmark/maskrcnn_benchmark/structures/segmentation_mask.py", line 393, in resize
    resized_polygons.append(polygon.resize(size))
  File "/home/ubuntu/github/maskrcnn-benchmark/maskrcnn_benchmark/structures/segmentation_mask.py", line 277, in resize
    ratios = tuple(float(s) / float(s_orig) for s, s_orig in zip(size, self.size))
  File "/home/ubuntu/github/maskrcnn-benchmark/maskrcnn_benchmark/structures/segmentation_mask.py", line 277, in <genexpr>
    ratios = tuple(float(s) / float(s_orig) for s, s_orig in zip(size, self.size))
ZeroDivisionError: float division by zero

Does anyone have any insight on this issue? Many thanks!

[johncorring]

Looks like cropping is leaving some boxes empty and so the masks/boxes have 0 area. Pretty sure you need to filter out empty/severely cropped boxes when you call bbox.crop.

[shamazharikh]

I used the original code and met the same problem

Did you solve it?

No, this problem randomly occurs and only occurred once, when I resume training and train for another backbone, seems that this problem is gone, sorry that I also have no idea why this happened.

I have encountered similar problem before that the randomly cropped path does not have any boxes so it cannot perform training normally. I have implemented a new version and it works good now.

`class RandomCrop(object):

Random crop data augmentation for training.

This is motivated from papers of face detection on WiderFace.

The min_scale and max_scale are relative to the smaller side

of the input image. First, we crop a square region by randomly

sample a float scale within [min_scale, max_scale], and then

crop the image. Second, the boxes are cropped accordingly with

boxes kept only if the IoU with original box is larger than

the given iou_thresh. If it ends up without any remaining boxes,

it will randomly select a box to re-crop to make sure the box is

within the cropped region (to guard the process to not crush

due to no target for an image).

def __init__(self, min_scale, max_scale, iou_thresh=0.5):
    # scale realtive to the smaller size of the image
    self.min_scale = min_scale
    self.max_scale = max_scale
    self.iou_thresh = iou_thresh

def get_params(self, image_size):
    w, h = image_size
    min_s = self.min_scale
    max_s = self.max_scale
    # randomly sample a scale within the range
    s = random.random() * (max_s - min_s) + min_s
    crop_size = int(min(w, h) * s)
    x1 = random.randint(0, w - crop_size)
    y1 = random.randint(0, h - crop_size)
    return x1, y1, crop_size

def get_safe_params(self, image_size, crop_size, target):
    w, h = image_size
    b = random.randint(0, len(target)-1) 
    x1,y1,x2,y2 = target[[b]]._split_into_xyxy()
    cx = int((x1 + x2) / 2)
    cy = int((y1 + y2) / 2)
    x1 = max(int(cx - crop_size/2), 0)
    y1 = max(int(cy - crop_size/2), 0)
    x2 = min(int(cx + crop_size/2), w)
    y2 = min(int(cy + crop_size/2), h)
    new_crop_size = min(x2-x1+1, y2-y1+1)
    return x1, y1, new_crop_size

def __call__(self, image, target):
    image_size = image.size
    x1, y1, crop_size = self.get_params(image_size)

    original_target = target.copy_with_fields(list(target.extra_fields.keys()))
    box = (x1, y1, x1+crop_size-1, y1+crop_size-1)
    target = target.crop(box)
    ious = target.area() / original_target.area()
    target = target[ious >= self.iou_thresh]

    if len(target) > 0:
        image = F.crop(image, y1, x1, crop_size, crop_size)
        return image, target

    # guard against no box available after crop
    x1, y1, crop_size = self.get_safe_params(image_size, crop_size, original_target)
    box = (x1, y1, x1+crop_size-1, y1+crop_size-1)
    target = original_target.crop(box) # re-crop original target
    image = F.crop(image, y1, x1, crop_size, crop_size)
    return image, target        

`

Hi, On using the code, I am getting the following assertion error. Can you help me out on this?

assert all(bbox.size == size for bbox in bboxes) AssertionError