Add down_sampling.py for generating LQ image from GT image. (open-mmlab#222)

* Add down_sampling.py for generating LQ image from GT image, which is required in LIIF.

* Add '__repr__' and test_down_sampling.py.

* Add docstring, rename parameter and change the function of resize.

* Fine-tuning code and docstring of RandomDownSampling class.

* Remove hardcode of bicubic and pillow.

Co-authored-by: 李尹硕 <SENSETIME\liyinshuo@cn0014004493l.domain.sensetime.com>

Author: Yshuo-Li

mmedit/datasets/pipelines/__init__.py

@@ -5,6 +5,7 @@
 from .compose import Compose
 from .crop import (Crop, CropAroundCenter, CropAroundFg, CropAroundUnknown,
                    FixedCrop, ModCrop, PairedRandomCrop)
+from .down_sampling import RandomDownSampling
 from .formating import (Collect, FormatTrimap, GetMaskedImage, ImageToTensor,
                         ToTensor)
 from .loading import (GetSpatialDiscountMask, LoadImageFromFile,
@@ -25,6 +26,6 @@
     'MergeFgAndBg', 'CompositeFg', 'TemporalReverse', 'LoadImageFromFileList',
     'GenerateFrameIndices', 'GenerateFrameIndiceswithPadding', 'FixedCrop',
     'LoadPairedImageFromFile', 'GenerateSoftSeg', 'GenerateSeg', 'PerturbBg',
-    'CropAroundFg', 'GetSpatialDiscountMask',
+    'CropAroundFg', 'GetSpatialDiscountMask', 'RandomDownSampling',
     'GenerateTrimapWithDistTransform', 'TransformTrimap'
 ]

mmedit/datasets/pipelines/down_sampling.py

@@ -0,0 +1,122 @@
+import math
+
+import numpy as np
+import torch
+from mmcv import imresize
+
+from ..registry import PIPELINES
+
+
+@PIPELINES.register_module()
+class RandomDownSampling:
+    """Generate LQ image from GT (and crop), which will randomly pick a scale.
+
+    Args:
+        scale_min (float): The minimum of upsampling scale, inclusive.
+            Default: 1.0.
+        scale_max (float): The maximum of upsampling scale, exclusive.
+            Default: 4.0.
+        patch_size (int): The cropped lr patch size.
+            Default: None, means no crop.
+        interpolation (str): Interpolation method, accepted values are
+            "nearest", "bilinear", "bicubic", "area", "lanczos" for 'cv2'
+            backend, "nearest", "bilinear", "bicubic", "box", "lanczos",
+            "hamming" for 'pillow' backend.
+            Default: "bicubic".
+        backend (str | None): The image resize backend type. Options are `cv2`,
+            `pillow`, `None`. If backend is None, the global imread_backend
+            specified by ``mmcv.use_backend()`` will be used.
+            Default: "pillow".
+
+        Scale will be picked in the range of [scale_min, scale_max).
+    """
+
+    def __init__(self,
+                 scale_min=1.0,
+                 scale_max=4.0,
+                 patch_size=None,
+                 interpolation='bicubic',
+                 backend='pillow'):
+        assert scale_max >= scale_min
+        self.scale_min = scale_min
+        self.scale_max = scale_max
+        self.patch_size = patch_size
+        self.interpolation = interpolation
+        self.backend = backend
+
+    def __call__(self, results):
+        """Call function.
+
+        Args:
+            results (dict): A dict containing the necessary information and
+                data for augmentation. 'gt' is required.
+
+        Returns:
+            dict: A dict containing the processed data and information.
+                modified 'gt', supplement 'lq' and 'scale' to keys.
+        """
+        img = results['gt']
+        scale = np.random.uniform(self.scale_min, self.scale_max)
+
+        if self.patch_size is None:
+            h_lr = math.floor(img.shape[-3] / scale + 1e-9)
+            w_lr = math.floor(img.shape[-2] / scale + 1e-9)
+            img = img[:round(h_lr * scale), :round(w_lr * scale), :]
+            img_down = resize_fn(img, (w_lr, h_lr), self.interpolation,
+                                 self.backend)
+            crop_lr, crop_hr = img_down, img
+        else:
+            w_lr = self.patch_size
+            w_hr = round(w_lr * scale)
+            x0 = np.random.randint(0, img.shape[-3] - w_hr)
+            y0 = np.random.randint(0, img.shape[-2] - w_hr)
+            crop_hr = img[x0:x0 + w_hr, y0:y0 + w_hr, :]
+            crop_lr = resize_fn(crop_hr, w_lr, self.interpolation,
+                                self.backend)
+        results['gt'] = crop_hr
+        results['lq'] = crop_lr
+        results['scale'] = scale
+
+        return results
+
+    def __repr__(self):
+        repr_str = self.__class__.__name__
+        repr_str += (f'scale_min={self.scale_min}, '
+                     f'scale_max={self.scale_max}, '
+                     f'patch_size={self.patch_size}')
+
+        return repr_str
+
+
+def resize_fn(img, size, interpolation='bicubic', backend='pillow'):
+    """Resize the given image to a given size.
+
+    Args:
+        img (ndarray | torch.Tensor): The input image.
+        size (int | tuple[int]): Target size w or (w, h).
+        interpolation (str): Interpolation method, accepted values are
+            "nearest", "bilinear", "bicubic", "area", "lanczos" for 'cv2'
+            backend, "nearest", "bilinear", "bicubic", "box", "lanczos",
+            "hamming" for 'pillow' backend.
+            Default: "bicubic".
+        backend (str | None): The image resize backend type. Options are `cv2`,
+            `pillow`, `None`. If backend is None, the global imread_backend
+            specified by ``mmcv.use_backend()`` will be used.
+            Default: "pillow".
+
+    Returns:
+        ndarray | torch.Tensor: `resized_img`, whose type is same as `img`.
+    """
+    if isinstance(size, int):
+        size = (size, size)
+    if isinstance(img, np.ndarray):
+        return imresize(
+            img, size, interpolation=interpolation, backend=backend)
+    elif isinstance(img, torch.Tensor):
+        image = imresize(
+            img.numpy(), size, interpolation=interpolation, backend=backend)
+        return torch.from_numpy(image)
+
+    else:
+        raise TypeError('img should got np.ndarray or torch.Tensor,'
+                        f'but got {type(img)}')

tests/test_down_sampling.py

@@ -0,0 +1,29 @@
+import numpy as np
+
+from mmedit.datasets.pipelines import RandomDownSampling
+
+
+def test_down_sampling():
+    img1 = np.uint8(np.random.randn(480, 640, 3) * 255)
+    inputs1 = dict(gt=img1)
+    down_sampling1 = RandomDownSampling(
+        scale_min=1, scale_max=4, patch_size=None)
+    results1 = down_sampling1(inputs1)
+    assert set(list(results1.keys())) == set(['gt', 'lq', 'scale'])
+    assert repr(down_sampling1) == (
+        down_sampling1.__class__.__name__ +
+        f'scale_min={down_sampling1.scale_min}, ' +
+        f'scale_max={down_sampling1.scale_max}, ' +
+        f'patch_size={down_sampling1.patch_size}')
+
+    img2 = np.uint8(np.random.randn(480, 640, 3) * 255)
+    inputs2 = dict(gt=img2)
+    down_sampling2 = RandomDownSampling(
+        scale_min=1, scale_max=4, patch_size=48)
+    results2 = down_sampling2(inputs2)
+    assert set(list(results2.keys())) == set(['gt', 'lq', 'scale'])
+    assert repr(down_sampling2) == (
+        down_sampling2.__class__.__name__ +
+        f'scale_min={down_sampling2.scale_min}, ' +
+        f'scale_max={down_sampling2.scale_max}, ' +
+        f'patch_size={down_sampling2.patch_size}')