Remove DPIAdjusting transform

dan/manager/dataset.py

@@ -347,19 +347,6 @@ def apply_preprocessing(sample, preprocessings):
     resize_ratio = [1, 1]
     img = sample["img"]
     for preprocessing in preprocessings:
-        if preprocessing["type"] == "dpi":
-            ratio = preprocessing["target"] / preprocessing["source"]
-            temp_img = img
-            h, w, c = temp_img.shape
-            temp_img = cv2.resize(
-                temp_img, (int(np.ceil(w * ratio)), int(np.ceil(h * ratio)))
-            )
-            if len(temp_img.shape) == 2:
-                temp_img = np.expand_dims(temp_img, axis=2)
-            img = temp_img
-
-            resize_ratio = [ratio, ratio]
-
         if preprocessing["type"] == "to_grayscaled":
             temp_img = img
             h, w, c = temp_img.shape

dan/transforms.py

@@ -20,22 +20,6 @@ from torchvision.transforms import (
 from torchvision.transforms.functional import InterpolationMode
 
 
-class DPIAdjusting:
-    """
-    Resolution modification
-    """
-
-    def __init__(self, factor):
-        self.factor = factor
-
-    def __call__(self, x):
-        w, h = x.size
-        return x.resize(
-            (int(np.ceil(w * self.factor)), int(np.ceil(h * self.factor))),
-            Image.BILINEAR,
-        )
-
-
 class Dilation:
     """
     OCR: stroke width increasing
@@ -175,33 +159,8 @@ def get_list_augmenters(img, aug_configs, fill_value):
     for aug_config in aug_configs:
         if rand((1,)) > aug_config["proba"]:
             continue
-        if aug_config["type"] == "dpi":
-            valid_factor = False
-            while not valid_factor:
-                factor = Uniform(
-                    aug_config["min_factor"], aug_config["max_factor"]
-                ).sample()
-                valid_factor = not (
-                    (
-                        "max_width" in aug_config
-                        and factor * img.size[0] > aug_config["max_width"]
-                    )
-                    or (
-                        "max_height" in aug_config
-                        and factor * img.size[1] > aug_config["max_height"]
-                    )
-                    or (
-                        "min_width" in aug_config
-                        and factor * img.size[0] < aug_config["min_width"]
-                    )
-                    or (
-                        "min_height" in aug_config
-                        and factor * img.size[1] < aug_config["min_height"]
-                    )
-                )
-            augmenters.append(DPIAdjusting(factor))
 
-        elif aug_config["type"] == "zoom_ratio":
+        if aug_config["type"] == "zoom_ratio":
             ratio_h = Uniform(
                 aug_config["min_ratio_h"], aug_config["max_ratio_h"]
             ).sample()
@@ -340,12 +299,6 @@ def aug_config(proba_use_da, p):
         "order": "random",
         "proba": proba_use_da,
         "augmentations": [
-            {
-                "type": "dpi",
-                "proba": p,
-                "min_factor": 0.75,
-                "max_factor": 1,
-            },
             {
                 "type": "perspective",
                 "proba": p,

docs/usage/train/parameters.md

@@ -23,14 +23,6 @@ All hyperparameters are specified and editable in the training scripts (meaning
 Preprocessing is applied before training the network (see `dan/manager/dataset.py`).
 The following transformations are implemented:
 
-* DPI adjustment
-```py
-    {
-        "type": "dpi",
-        "source": 300,
-        "target": 150,
-    }
-```
 * Convert to grayscale
 ```py
 {
@@ -80,7 +72,6 @@ Augmentation transformations are applied on-the-fly during training to artificia
 
 The following transformations are implemented in `dan/transforms.py`:
 * Color inversion
-* DPI adjusting
 * Dilation and erosion
 * Elastic distortion
 * Reducing interline spacing
@@ -100,13 +91,6 @@ The following configuration is used by default when using the `teklia-dan train
         "order": "random",
         "proba": 0.9,
         "augmentations": [
-            {
-                "type": "dpi",
-                "proba": 0.1,
-                "min_factor": 0.75,
-                "max_factor": 1,
-                "preserve_ratio": True,
-            },
             {
                 "type": "perspective",
                 "proba": 0.1,

tests/test_training.py

@@ -16,7 +16,7 @@ from tests.conftest import FIXTURES
             "last_3.pt",
             {
                 "nb_chars": 43,
-                "cer": 1.2791,
+                "cer": 1.2558,
                 "nb_words": 9,
                 "wer": 1.0,
                 "nb_words_no_punct": 9,
@@ -34,7 +34,7 @@ from tests.conftest import FIXTURES
             },
             {
                 "nb_chars": 49,
-                "cer": 1.1429,
+                "cer": 1.1224,
                 "nb_words": 9,
                 "wer": 1.0,
                 "nb_words_no_punct": 9,
@@ -99,7 +99,7 @@ def test_train_and_test(
                     assert torch.allclose(
                         trained_tensor,
                         expected_tensor,
-                        atol=1e-04,
+                        atol=1e-03,
                     )
             assert (
                 trained_model[optimizer_part]["param_groups"]