evalscope 1.0.0__py3-none-any.whl → 1.2.0__py3-none-any.whl

evalscope/benchmarks/ocr_bench/ocr_bench/ocr_bench_adapter.py

@@ -0,0 +1,101 @@
+import json
+from typing import Any, Dict, List
+
+from evalscope.api.benchmark import BenchmarkMeta, VisionLanguageAdapter
+from evalscope.api.dataset import Sample
+from evalscope.api.evaluator.state import TaskState
+from evalscope.api.messages import ChatMessageUser, Content, ContentImage, ContentText
+from evalscope.api.metric.scorer import Score
+from evalscope.api.registry import register_benchmark
+from evalscope.constants import Tags
+from evalscope.utils.io_utils import bytes_to_base64
+from evalscope.utils.logger import get_logger
+
+logger = get_logger()
+
+SUBSET_LIST = [
+    'Regular Text Recognition', 'Irregular Text Recognition', 'Artistic Text Recognition', 'Handwriting Recognition',
+    'Digit String Recognition', 'Non-Semantic Text Recognition', 'Scene Text-centric VQA', 'Doc-oriented VQA',
+    'Key Information Extraction', 'Handwritten Mathematical Expression Recognition'
+]
+
+
+@register_benchmark(
+    BenchmarkMeta(
+        name='ocr_bench',
+        pretty_name='OCRBench',
+        tags=[Tags.MULTI_MODAL, Tags.KNOWLEDGE, Tags.QA],
+        description=
+        'OCRBench is a comprehensive evaluation benchmark designed to assess the OCR capabilities of Large Multimodal Models. It comprises five components: Text Recognition, SceneText-Centric VQA, Document-Oriented VQA, Key Information Extraction, and Handwritten Mathematical Expression Recognition. The benchmark includes 1000 question-answer pairs, and all the answers undergo manual verification and correction to ensure a more precise evaluation.',  # noqa: E501
+        dataset_id='evalscope/OCRBench',
+        subset_list=SUBSET_LIST,
+        metric_list=['acc'],
+        eval_split='test',
+        prompt_template='{question}',
+    )
+)
+class OCRBenchAdapter(VisionLanguageAdapter):
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self.add_aggregation_name = False
+        self.reformat_subset = True
+
+    def record_to_sample(self, record: Dict[str, Any]) -> Sample:
+
+        input_text = self.prompt_template.format(question=record['question'])
+        content_list: List[Content] = [ContentText(text=input_text)]
+        image = record.get('image')
+        if image:
+            image_base64 = bytes_to_base64(image['bytes'], format='jpeg', add_header=True)
+            content_list.append(ContentImage(image=image_base64))
+        return Sample(
+            input=[ChatMessageUser(content=content_list)],
+            target=json.dumps(record.get('answer'), ensure_ascii=False),  # answers is a list
+            subset_key=record.get('question_type'),
+            metadata={
+                'dataset': record.get('dataset'),
+                'question_type': record.get('question_type'),
+            }
+        )
+
+    def match_score(
+        self, original_prediction: str, filtered_prediction: str, reference: str, task_state: TaskState
+    ) -> Score:
+
+        score = Score(
+            extracted_prediction=filtered_prediction,
+            prediction=original_prediction,
+        )
+
+        pred = filtered_prediction.lower().strip()
+        gt_ans = json.loads(reference)
+        dataset_name = task_state.metadata['dataset']
+
+        score_value = 0
+        if dataset_name == 'HME100k':
+            if isinstance(gt_ans, list):
+                for j in range(len(gt_ans)):
+                    answer = gt_ans[j].strip().replace('\n', ' ').replace(' ', '')
+                    predict = pred.strip().replace('\n', ' ').replace(' ', '')
+                    if answer in predict:
+                        score_value = 1
+            else:
+                answer = gt_ans.strip().replace('\n', ' ').replace(' ', '')
+                predict = pred.strip().replace('\n', ' ').replace(' ', '')
+                if answer in predict:
+                    score_value = 1
+        else:
+            if isinstance(gt_ans, list):
+                for j in range(len(gt_ans)):
+                    answer = gt_ans[j].lower().strip().replace('\n', ' ')
+                    predict = pred.lower().strip().replace('\n', ' ')
+                    if answer in predict:
+                        score_value = 1
+            else:
+                answer = gt_ans.lower().strip().replace('\n', ' ')
+                predict = pred.lower().strip().replace('\n', ' ')
+                if answer in predict:
+                    score_value = 1
+        score.value = {'acc': score_value}
+        return score

evalscope/benchmarks/ocr_bench/ocr_bench_v2/IoUscore_metric.py

@@ -0,0 +1,87 @@
+# flake8: noqa
+import ast
+import re
+
+from .vqa_metric import vqa_evaluation
+
+
+def calculate_iou(box1, box2):
+    try:
+        box1 = [int(coordinate) for coordinate in box1]
+        box2 = [int(coordinate) for coordinate in box2]
+    except:
+        return 0
+
+    x1_inter = max(box1[0], box2[0])
+    y1_inter = max(box1[1], box2[1])
+    x2_inter = min(box1[2], box2[2])
+    y2_inter = min(box1[3], box2[3])
+
+    inter_area = max(0, x2_inter - x1_inter) * max(0, y2_inter - y1_inter)
+
+    box1_area = (box1[2] - box1[0]) * (box1[3] - box1[1])
+    box2_area = (box2[2] - box2[0]) * (box2[3] - box2[1])
+
+    union_area = box1_area + box2_area - inter_area
+
+    iou = inter_area / union_area if union_area != 0 else 0
+
+    return iou
+
+
+def vqa_with_position_evaluation(predict, img_metas):
+    score_content, score_bbox = 0.0, 0.0
+    if 'answer' in predict.keys():
+        score_content = vqa_evaluation(predict['answer'], img_metas['answers'])
+    if 'bbox' in predict.keys():
+        gt_bbox = img_metas['bbox']
+        try:
+            predict_bbox_list = ast.literal_eval(predict['bbox'])
+            score_bbox = calculate_iou(predict_bbox_list, gt_bbox)
+        except:
+            score_bbox = 0
+    return 0.5 * score_content + 0.5 * score_bbox
+
+
+def extract_coordinates(text):
+    # Regex pattern to match coordinates in either (x1, y1, x2, y2) or [x1, y1, x2, y2] format
+
+    pattern = r'[\(\[]\s*(\d+)\s*,\s*(\d+)\s*,\s*(\d+)\s*,\s*(\d+)\s*[\)\]]'
+
+    matches = list(re.finditer(pattern, text))
+    coords_list = []
+    coords_set = set()
+    for match in matches:
+        x1, y1, x2, y2 = map(int, match.groups())
+
+        if all(0 <= n <= 1000 for n in [x1, y1, x2, y2]):
+            coords = (x1, y1, x2, y2)
+
+            if coords in coords_set:
+                coords_list = [c for c in coords_list if c != coords]
+
+            coords_list.append(coords)
+            coords_set.add(coords)
+    if coords_list:
+        last_coords = coords_list[-1]
+        return list(last_coords)
+    else:
+        return None
+
+
+if __name__ == '__main__':
+    print('Example for Text Grounding task.')
+    box1 = [50, 50, 150, 150]
+    box2 = [60, 60, 140, 140]
+    iou_score = calculate_iou(box1, box2)
+    print(f'IoU score: {iou_score}')
+
+    print('Example for VQA with position task.')
+    pred = {'content': 'The content is Hello Buddies', 'bbox': box1}
+    gt = {'content': 'Hello Buddies', 'bbox': box2}
+
+    vqa_score = vqa_evaluation(pred['content'], gt['content'])
+    iou_score = calculate_iou(pred['bbox'], gt['bbox'])
+
+    print(f'VQA score: {vqa_score}')
+    print(f'IoU score: {iou_score}')