Detection module : Preprocessor

doctr/models/__init__.py

@@ -1 +1,2 @@
 from . import utils
+from . import preprocessor

doctr/models/preprocessor.py

@@ -0,0 +1,93 @@
+# Copyright (C) 2021, Mindee.
+# This program is licensed under the Apache License version 2.
+# See LICENSE or go to <https://www.apache.org/licenses/LICENSE-2.0.txt> for full license details.
+
+import math
+import cv2
+import json
+import os
+import numpy as np
+from typing import Union, List, Tuple, Optional, Any, Dict
+
+
+class Preprocessor():
+
+    def __init__(
+        self,
+        out_size: Tuple[int, int],
+        normalization: bool = True,
+        batch_size: int = 1
+    ) -> None:
+
+        self.out_size = out_size
+        self.normalization = normalization
+        self.batch_size = batch_size
+
+    def __call__(
+        documents: Tuple[List[List[np.ndarray]], List[List[str]], List[List[Tuple[int, int]]]]
+    ) -> List[Tuple[List[np.ndarray], List[str], List[Tuple[int, int]]]]:
+        """
+        perform resizing, normalization and batching on documents
+        """
+        images, names, shapes = documents
+        images = resize_documents_imgs(images, out_size=self.out_size)
+        if self.normalization:
+            images = normalize_documents_imgs(images)
+        norm_and_sized_docs = images, names, shapes
+        b_docs, docs_indexes, pages_indexes = batch_documents(
+            norm_and_sized_docs, batch_size=self.batch_size)
+
+        return b_docs, docs_indexes, pages_indexes
+
+    def normalize_documents_imgs(
+        documents_imgs: List[List[np.ndarray]],
+        mode: str = 'symmetric'
+    ) -> List[List[np.ndarray]]:
+        """
+        normalize documents imgs according to mode
+        """
+
+        if mode == 'symmetric':
+            return [[(img - 128) / 128 for img in doc] for doc in documents_imgs]
+
+    def resize_documents_imgs(
+        documents_imgs: List[List[np.ndarray]],
+        out_size: Tuple[int, int]
+    ) -> List[List[np.ndarray]]:
+        """
+        Resize documents img to the out_size : size for the model inputs
+        The nested structure documents/pages is preserved
+        returns resized documents img
+        """
+        return [[cv2.resize(img, out_size, cv2.INTER_LINEAR) for img in doc] for doc in documents_imgs]
+
+    def batch_documents(
+        documents: Tuple[List[List[np.ndarray]], List[List[str]], List[List[Tuple[int, int]]]],
+        batch_size: int = 1
+    ) -> List[Tuple[List[np.ndarray], List[str], List[Tuple[int, int]]]]:
+        """
+        function to batch a list of read documents
+        :param documents: documents read by documents.reader.read_documents
+        :param batch_size: batch_size to use during inference, default goes to 1
+        """
+
+        images, names, shapes = documents
+
+        # keep track of both documents and pages indexes
+        docs_indexes = [images.index(doc) for doc in images for _ in doc]
+        pages_indexes = [doc.index(page) for doc in images for page in doc]
+
+        # flatten structure
+        flat_images = [image for doc in images for raw in doc]
+        flat_names = [name for doc in names for name in doc]
+        flat_shapes = [shape for doc in shapes for shape in doc]
+
+        range_batch = range((len(flat_shapes) + batch_size - 1) // batch_size)
+
+        b_images = [flat_images[i * batch_size:(i + 1) * batch_size] for i in range_batch]
+        b_names = [flat_names[i * batch_size:(i + 1) * batch_size] for i in range_batch]
+        b_shapes = [flat_shapes[i * batch_size:(i + 1) * batch_size] for i in range_batch]
+
+        b_docs = [[b_i, b_n, b_s] for b_i, b_n, b_s in zip(b_images, b_names, b_shapes)]
+
+        return b_docs, docs_indexes, pages_indexes

test/test_preprocessor.py

@@ -0,0 +1,27 @@
+import pytest
+import requests
+from io import BytesIO
+
+from doctr.models.preprocessor import Preprocessor
+from doctr import documents
+
+@pytest.fixture(scope="session")
+def mock_pdf(tmpdir_factory):
+    url = 'https://arxiv.org/pdf/1911.08947.pdf'
+    file = BytesIO(requests.get(url).content)
+    fn = tmpdir_factory.mktemp("data").join("mock_pdf_file.pdf")
+    with open(fn, 'wb') as f:
+        f.write(file.getbuffer())
+    return fn
+
+
+def test_preprocess_documents(num_docs=10, batch_size=3):
+    docs = documents.reader.read_documents(
+        filepaths=[mock_pdf for _ in range(num_docs)])
+    preprocessor = Preprocessor(out_size=(600, 600), normalization=True, batch_size=batch_size)
+    batched_docs, docs_indexes, pages_indexes = preprocessor(docs)
+    assert len(docs_indexes) == len(pages_indexes)
+    if num_docs > batch_size:
+        for batch in batched_docs[:-1]:
+            for i in range(len(batch)):
+                assert len(batch[i]) == batch_size