Resize the input image so that its longest side and shortest side are within a specified range, ensuring that both sides are divisible by a specified stride. Args: max_size (int): Maximum size for the longest edge of the image. min_size (int): Minimum size for the shortest e
| 32 | from modeling.qwen2vl.image_processing_qwen2_vl import Qwen2VLImageProcessor |
| 33 | |
| 34 | class MaxLongEdgeMinShortEdgeResize(torch.nn.Module): |
| 35 | """Resize the input image so that its longest side and shortest side are within a specified range, |
| 36 | ensuring that both sides are divisible by a specified stride. |
| 37 | |
| 38 | Args: |
| 39 | max_size (int): Maximum size for the longest edge of the image. |
| 40 | min_size (int): Minimum size for the shortest edge of the image. |
| 41 | stride (int): Value by which the height and width of the image must be divisible. |
| 42 | max_pixels (int): Maximum pixels for the full image. |
| 43 | interpolation (InterpolationMode): Desired interpolation enum defined by |
| 44 | :class:`torchvision.transforms.InterpolationMode`. Default is ``InterpolationMode.BILINEAR``. |
| 45 | If input is Tensor, only ``InterpolationMode.NEAREST``, ``InterpolationMode.NEAREST_EXACT``, |
| 46 | ``InterpolationMode.BILINEAR``, and ``InterpolationMode.BICUBIC`` are supported. |
| 47 | The corresponding Pillow integer constants, e.g., ``PIL.Image.BILINEAR`` are also accepted. |
| 48 | antialias (bool, optional): Whether to apply antialiasing (default is True). |
| 49 | """ |
| 50 | |
| 51 | def __init__( |
| 52 | self, |
| 53 | max_size: int, |
| 54 | min_size: int, |
| 55 | stride: int, |
| 56 | max_pixels: int, |
| 57 | interpolation=InterpolationMode.BICUBIC, |
| 58 | antialias=True |
| 59 | ): |
| 60 | super().__init__() |
| 61 | self.max_size = max_size |
| 62 | self.min_size = min_size |
| 63 | self.stride = stride |
| 64 | self.max_pixels = max_pixels |
| 65 | self.interpolation = interpolation |
| 66 | self.antialias = antialias |
| 67 | |
| 68 | def _make_divisible(self, value, stride): |
| 69 | """Ensure the value is divisible by the stride.""" |
| 70 | return max(stride, int(round(value / stride) * stride)) |
| 71 | |
| 72 | def _apply_scale(self, width, height, scale): |
| 73 | new_width = round(width * scale) |
| 74 | new_height = round(height * scale) |
| 75 | new_width = self._make_divisible(new_width, self.stride) |
| 76 | new_height = self._make_divisible(new_height, self.stride) |
| 77 | return new_width, new_height |
| 78 | |
| 79 | def forward(self, img, img_num=1): |
| 80 | """ |
| 81 | Args: |
| 82 | img (PIL Image): Image to be resized. |
| 83 | img_num (int): Number of images, used to change max_tokens. |
| 84 | Returns: |
| 85 | PIL Image or Tensor: Rescaled image with divisible dimensions. |
| 86 | """ |
| 87 | if isinstance(img, torch.Tensor): |
| 88 | height, width = img.shape[-2:] |
| 89 | else: |
| 90 | width, height = img.size |
| 91 |