No more DPI adjusting

Closes #191 (closed)

Needs testing on a real dataset because training test is failing only because optimizers are different.

dan/ocr/transforms.py

@@ -15,16 +15,16 @@ from albumentations.augmentations import (
     GaussianBlur,
     GaussNoise,
     Perspective,
+    RandomScale,
     Sharpen,
     ToGray,
 )
 from albumentations.core.transforms_interface import ImageOnlyTransform
-from cv2 import dilate, erode, resize
+from cv2 import dilate, erode
 from numpy import random
 from torch import Tensor
-from torch.distributions.uniform import Uniform
 from torchvision.transforms import Compose, ToPILImage
-from torchvision.transforms.functional import resize as resize_tensor
+from torchvision.transforms.functional import resize
 
 
 class Preprocessing(str, Enum):
@@ -54,7 +54,7 @@ class FixedHeightResize:
 
     def __call__(self, img: Tensor) -> Tensor:
         size = (self.height, self._calc_new_width(img))
-        return resize_tensor(img, size, antialias=False)
+        return resize(img, size, antialias=False)
 
     def _calc_new_width(self, img: Tensor) -> int:
         aspect_ratio = img.shape[2] / img.shape[1]
@@ -71,7 +71,7 @@ class FixedWidthResize:
 
     def __call__(self, img: Tensor) -> Tensor:
         size = (self._calc_new_height(img), self.width)
-        return resize_tensor(img, size, antialias=False)
+        return resize(img, size, antialias=False)
 
     def _calc_new_height(self, img: Tensor) -> int:
         aspect_ratio = img.shape[1] / img.shape[2]
@@ -96,7 +96,7 @@ class MaxResize:
         ratio = min(height_ratio, width_ratio)
         new_width = int(width * ratio)
         new_height = int(height * ratio)
-        return resize_tensor(img, (new_height, new_width), antialias=False)
+        return resize(img, (new_height, new_width), antialias=False)
 
 
 class Dilation:
@@ -156,29 +156,6 @@ class ErosionDilation(ImageOnlyTransform):
         )
 
 
-class DPIAdjusting(ImageOnlyTransform):
-    """
-    Resolution modification
-    """
-
-    def __init__(
-        self,
-        min_factor: float = 0.75,
-        max_factor: float = 1,
-        always_apply: bool = False,
-        p: float = 1.0,
-    ):
-        super(DPIAdjusting, self).__init__(always_apply, p)
-        self.min_factor = min_factor
-        self.max_factor = max_factor
-        self.p = p
-        self.always_apply = False
-
-    def apply(self, img: np.ndarray, **params):
-        factor = float(Uniform(self.min_factor, self.max_factor).sample())
-        return resize(img, None, fx=factor, fy=factor)
-
-
 def get_preprocessing_transforms(
     preprocessings: list, to_pil_image: bool = False
 ) -> Compose:
@@ -212,22 +189,31 @@ def get_augmentation_transforms() -> A.Compose:
     """
     return A.Compose(
         [
-            DPIAdjusting(min_factor=0.75, max_factor=1),
+            # Scale between 0.75 and 1.0
+            RandomScale(scale_limit=[-0.25, 0], always_apply=True),
             A.SomeOf(
                 [
                     ErosionDilation(min_kernel=1, max_kernel=4, iterations=1),
                     Perspective(scale=(0.05, 0.09), fit_output=True, p=0.4),
-                    GaussianBlur(sigma_limit=2.5, p=1),
-                    GaussNoise(var_limit=50**2, p=1),
+                    GaussianBlur(sigma_limit=2.5, always_apply=True),
+                    GaussNoise(var_limit=50**2, always_apply=True),
                     ColorJitter(
-                        contrast=0.2, brightness=0.2, saturation=0.2, hue=0.2, p=1
+                        contrast=0.2,
+                        brightness=0.2,
+                        saturation=0.2,
+                        hue=0.2,
+                        always_apply=True,
                     ),
                     ElasticTransform(
-                        alpha=20.0, sigma=5.0, alpha_affine=1.0, border_mode=0, p=1
+                        alpha=20.0,
+                        sigma=5.0,
+                        alpha_affine=1.0,
+                        border_mode=0,
+                        always_apply=True,
                     ),
-                    Sharpen(alpha=(0.0, 1.0), p=1),
-                    Affine(shear={"x": (-20, 20), "y": (0, 0)}, p=1),
-                    CoarseDropout(p=1),
+                    Sharpen(alpha=(0.0, 1.0), always_apply=True),
+                    Affine(shear={"x": (-20, 20), "y": (0, 0)}, always_apply=True),
+                    CoarseDropout(always_apply=True),
                     ToGray(p=0.5),
                 ],
                 n=2,