Fix version 0.2.0-dev3 and later

In this MR, version 0.2.0-dev3 (and later) is fixed:

• We use DPIAdjusting transform as it is a key component when training on images of very different sizes
• We compute the real mean and std values on the train set (we do not use default values from ImageNet = revert a5f7b5ab)
• The padding is on the right and bottom of the image (= revert part of f57bc667 for now)
• Two parameters files are created during training: one for training details (used to reproduce/continue the experiment), one for inference (that can be directly used by Arkindex DAN worker)

dan/manager/dataset.py

@@ -22,18 +22,20 @@ class OCRDataset(Dataset):
         charset,
         tokens,
         preprocessing_transforms,
-        normalization_transforms,
         augmentation_transforms,
         load_in_memory=False,
+        mean=None,
+        std=None,
     ):
         self.set_name = set_name
         self.charset = charset
         self.tokens = tokens
         self.load_in_memory = load_in_memory
+        self.mean = mean
+        self.std = std
 
-        # Pre-processing, augmentation, normalization
+        # Pre-processing, augmentation
         self.preprocessing_transforms = preprocessing_transforms
-        self.normalization_transforms = normalization_transforms
         self.augmentation_transforms = augmentation_transforms
 
         # Factor to reduce the height and width of the feature vector before feeding the decoder.
@@ -55,20 +57,20 @@ class OCRDataset(Dataset):
         """
         Return an item from the dataset (image and label)
         """
-
         # Load preprocessed image
-        sample = dict(**self.samples[idx])
+        sample = copy.deepcopy(self.samples[idx])
         if not self.load_in_memory:
             sample["img"] = self.get_sample_img(idx)
 
+        # Convert to numpy
+        sample["img"] = np.array(sample["img"])
+
         # Apply data augmentation
         if self.augmentation_transforms:
-            sample["img"] = self.augmentation_transforms(image=np.array(sample["img"]))[
-                "image"
-            ]
+            sample["img"] = self.augmentation_transforms(image=sample["img"])["image"]
 
         # Image normalization
-        sample["img"] = self.normalization_transforms(sample["img"])
+        sample["img"] = (sample["img"] - self.mean) / self.std
 
         # Get final height and width
         sample["img_reduced_shape"], sample["img_position"] = self.compute_final_size(
@@ -113,28 +115,51 @@ class OCRDataset(Dataset):
 
     def get_sample_img(self, i):
         """
-        Compute the final image size and position after feature extraction
+        Get image by index
         """
         if self.load_in_memory:
             return self.samples[i]["img"]
 
         return self.preprocessing_transforms(read_image(self.samples[i]["path"]))
 
+    def compute_std_mean(self):
+        """
+        Compute cumulated variance and mean of whole dataset
+        """
+        if self.mean is not None and self.std is not None:
+            return self.mean, self.std
+
+        sum = np.zeros((3,))
+        diff = np.zeros((3,))
+        nb_pixels = 0
+        for metric in ["mean", "std"]:
+            for ind in range(len(self.samples)):
+                img = np.array(self.get_sample_img(ind))
+                if metric == "mean":
+                    sum += np.sum(img, axis=(0, 1))
+                    nb_pixels += np.prod(img.shape[:2])
+                elif metric == "std":
+                    diff += [
+                        np.sum((img[:, :, k] - self.mean[k]) ** 2) for k in range(3)
+                    ]
+            if metric == "mean":
+                self.mean = sum / nb_pixels
+            elif metric == "std":
+                self.std = np.sqrt(diff / nb_pixels)
+        return self.mean, self.std
+
     def compute_final_size(self, img):
         """
         Compute the final image size and position after feature extraction
         """
-        final_c, final_h, final_w = img.shape
-        image_reduced_shape = np.ceil(
-            [final_h, final_w, final_c] / self.reduce_dims_factor
-        ).astype(int)
+        image_reduced_shape = np.ceil(img.shape / self.reduce_dims_factor).astype(int)
 
         if self.set_name == "train":
             image_reduced_shape = [max(1, t) for t in image_reduced_shape]
 
         image_position = [
-            [0, final_h],
-            [0, final_w],
+            [0, img.shape[0]],
+            [0, img.shape[1]],
         ]
         return image_reduced_shape, image_position