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/ocr.py

@@ -244,7 +244,10 @@ class OCRCollateFunction:
         labels = [batch_data[i]["token_label"] for i in range(len(batch_data))]
         labels = pad_sequences_1D(labels, padding_value=self.label_padding_value).long()
 
-        imgs = [batch_data[i]["img"] for i in range(len(batch_data))]
+        imgs = [
+            torch.from_numpy(batch_data[i]["img"]).permute(2, 0, 1)
+            for i in range(len(batch_data))
+        ]
         imgs = pad_images(imgs)
 
         formatted_batch_data = {