Feature/phrase grounding recall filter (#139)

* + add phrase_grounding_recall_filter op

* * update new OP
+ make llava_to_dj and dj_to_llava support only_keep_caption mode

* + Add unit test for phrase_grounding

* + remove hf model caches automatically for unittest

* * download required nltk data when initializing the phrase_grounding_recall_filter

* * output the cleaning log when the cleaning actually happens

* * update Operator docs

* * fix some typos

* * removing hf models automatically after unit test is finished for clip and blip

Author: HYLcool

configs/config_all.yaml

@@ -5,10 +5,11 @@
                image_shape_filter, image_size_filter,
                image_text_matching_filter, image_text_similarity_filter,
                language_id_score_filter, maximum_line_length_filter,
-               perplexity_filter, special_characters_filter,
-               specified_field_filter, specified_numeric_field_filter,
-               stopwords_filter, suffix_filter, text_action_filter,
-               text_entity_dependency_filter, text_length_filter,
-               token_num_filter, word_num_filter, word_repetition_filter)
+               perplexity_filter, phrase_grounding_recall_filter,
+               special_characters_filter, specified_field_filter,
+               specified_numeric_field_filter, stopwords_filter, suffix_filter,
+               text_action_filter, text_entity_dependency_filter,
+               text_length_filter, token_num_filter, word_num_filter,
+               word_repetition_filter)
 
 # yapf: enable

data_juicer/ops/filter/__init__.py

@@ -0,0 +1,286 @@
+from typing import List
+
+import numpy as np
+from jsonargparse.typing import ClosedUnitInterval
+from loguru import logger
+from PIL import ImageOps
+
+from data_juicer.utils.availability_utils import AvailabilityChecking
+from data_juicer.utils.constant import Fields, StatsKeys
+from data_juicer.utils.mm_utils import (SpecialTokens, iou, load_image,
+                                        remove_special_tokens)
+from data_juicer.utils.model_utils import get_model, prepare_model
+
+from ..base_op import OPERATORS, Filter
+from ..op_fusion import LOADED_IMAGES
+
+OP_NAME = 'phrase_grounding_recall_filter'
+
+with AvailabilityChecking(['torch', 'transformers', 'nltk'], OP_NAME):
+
+    import torch
+    import transformers  # noqa: F401
+
+    # avoid hanging when calling clip in multiprocessing
+    torch.set_num_threads(1)
+
+    import nltk
+
+
+# NER algorithm adapted from GLIP starts
+# https://github.com/microsoft/GLIP/blob/main/maskrcnn_benchmark/engine/predictor_glip.py#L107-L127
+def find_noun_phrases(caption: str) -> List[str]:
+    caption = caption.lower()
+    tokens = nltk.word_tokenize(caption)
+    pos_tags = nltk.pos_tag(tokens)
+
+    grammar = 'NP: {<DT>?<JJ.*>*<NN.*>+}'
+    cp = nltk.RegexpParser(grammar)
+    result = cp.parse(pos_tags)
+
+    noun_phrases = list()
+    for subtree in result.subtrees():
+        if subtree.label() == 'NP':
+            noun_phrases.append(' '.join(t[0] for t in subtree.leaves()))
+
+    return noun_phrases
+
+
+def remove_punctuation(text: str) -> str:
+    punct = [
+        '|', ':', ';', '@', '(', ')', '[', ']', '{', '}', '^', '\'', '\"', '’',
+        '`', '?', '$', '%', '#', '!', '&', '*', '+', ',', '.'
+    ]
+    for p in punct:
+        text = text.replace(p, '')
+    return text.strip()
+
+
+def run_ner(caption):
+    noun_phrases = find_noun_phrases(caption)
+    noun_phrases = [remove_punctuation(phrase) for phrase in noun_phrases]
+    noun_phrases = [phrase for phrase in noun_phrases if phrase != '']
+    noun_phrases = list(set(noun_phrases))  # remove duplicate ners
+    return noun_phrases
+
+
+# NER algorithm adapted from GLIP ends
+
+
+@OPERATORS.register_module(OP_NAME)
+@LOADED_IMAGES.register_module(OP_NAME)
+class PhraseGroundingRecallFilter(Filter):
+    """Filter to keep samples whose locating recalls of phrases extracted
+    from text in the images are within a specified range."""
+
+    def __init__(self,
+                 hf_owlvit='google/owlvit-base-patch32',
+                 min_recall: ClosedUnitInterval = 0.1,
+                 max_recall: ClosedUnitInterval = 1.0,
+                 horizontal_flip: bool = False,
+                 vertical_flip: bool = False,
+                 any_or_all: str = 'any',
+                 reduce_mode: str = 'avg',
+                 iou_thr: ClosedUnitInterval = 0.5,
+                 large_area_ratio_thr: ClosedUnitInterval = 0.95,
+                 conf_thr: ClosedUnitInterval = 0.0,
+                 *args,
+                 **kwargs):
+        """
+        Initialization method.
+
+        :param hf_owlvit: Owl-ViT model name on huggingface to locate the
+            phrases extracted from the text.
+        :param min_recall: The min phrase grounding recall to keep samples.
+        :param max_recall: The max phrase grounding recall to keep samples.
+        :param horizontal_flip: Flip image horizontally (left to right).
+        :param vertical_flip: Flip image vertically (top to bottom).
+        :param any_or_all: keep this sample with 'any' or 'all' strategy of
+            all images. 'any': keep this sample if any images meet the
+            condition. 'all': keep this sample only if all images meet the
+            condition.
+        :param reduce_mode: reduce mode when one text corresponds to
+            multiple images in a chunk.
+            'avg': Take the average of multiple values
+            'max': Take the max of multiple values
+            'min': Take the min of multiple values
+        :param iou_thr: the IoU threshold for NMS-like post-process. If two
+            predicted bboxes are overlap with an IoU larger than this
+            threshold, the bbox with less confidence will be removed. Default:
+            0.5.
+        :param large_area_ratio_thr: the area ratio threshold for filtering out
+            those large predicted bboxes. If the area of a predicted bbox
+            accounts for more than this ratio threshold of the whole image
+            area, this bbox will be removed. Default: 0.95.
+        :param conf_thr: the confidence score threshold for removing
+            low-confidence bboxes. If the confidence score of a predicted bbox
+            is lower than the threshold, this bbox will be removed. Default: 0.
+        :param args: extra args
+        :param kwargs: extra args
+        """
+        super().__init__(*args, **kwargs)
+        self.min_recall = min_recall
+        self.max_recall = max_recall
+        if reduce_mode not in ['avg', 'max', 'min']:
+            raise ValueError(f'Reduce mode [{reduce_mode}] is not supported. '
+                             f'Can only be one of ["avg", "max", "min"].')
+        if any_or_all not in ['any', 'all']:
+            raise ValueError(f'Keep strategy [{any_or_all}] is not supported. '
+                             f'Can only be one of ["any", "all"].')
+        self.any = (any_or_all == 'any')
+        self.model_type = 'hf_owlvit'
+        self.model_key = prepare_model(model_type=self.model_type,
+                                       model_key=hf_owlvit)
+        self.reduce_mode = reduce_mode
+        self.horizontal_flip = horizontal_flip
+        self.vertical_flip = vertical_flip
+
+        self.iou_thr = iou_thr
+        self.large_area_ratio_thr = large_area_ratio_thr
+        self.conf_thr = conf_thr
+
+        requires_nltk_data = ['punkt', 'averaged_perceptron_tagger']
+        logger.info(f'Downloading nltk data of {requires_nltk_data}...')
+        for nltk_data_pkg in requires_nltk_data:
+            nltk.download(nltk_data_pkg)
+
+    def compute_stats(self, sample, context=False):
+        # check if it's computed already
+        if StatsKeys.phrase_grounding_recall in sample[Fields.stats]:
+            return sample
+
+        # there is no image in this sample
+        if self.image_key not in sample or not sample[self.image_key]:
+            sample[Fields.stats][StatsKeys.phrase_grounding_recall] = np.array(
+                [], dtype=np.float64)
+            return sample
+
+        # load images
+        loaded_image_keys = sample[self.image_key]
+        images = {}
+        for loaded_image_key in loaded_image_keys:
+            if context and loaded_image_key in sample[Fields.context]:
+                # load from context
+                images[loaded_image_key] = sample[
+                    Fields.context][loaded_image_key]
+            else:
+                if loaded_image_key not in images:
+                    # avoid load the same images
+                    image = load_image(loaded_image_key)
+                    images[loaded_image_key] = image
+                    if context:
+                        # store the image data into context
+                        sample[Fields.context][loaded_image_key] = image
+
+        text = sample[self.text_key]
+        offset = 0
+        recalls = []
+        model, processor = get_model(self.model_key,
+                                     model_type=self.model_type)
+
+        for chunk in text.split(SpecialTokens.eoc):
+            count = chunk.count(SpecialTokens.image)
+
+            # no image or no text
+            if count == 0 or len(chunk) == 0:
+                continue
+            else:
+                text_this_chunk = remove_special_tokens(chunk)
+                ners_this_chunk = run_ner(text_this_chunk)
+                num_ners = len(ners_this_chunk)
+                if num_ners <= 0:
+                    # no ners found, just skip this chunk
+                    recalls.append(1.0)
+                    continue
+                images_this_chunk = []
+                for image_key in loaded_image_keys[offset:offset + count]:
+                    image = images[image_key]
+                    if self.horizontal_flip:
+                        image = ImageOps.mirror(image)
+                    if self.vertical_flip:
+                        image = ImageOps.flip(image)
+                    images_this_chunk.append(image)
+
+                ners_batch = [ners_this_chunk] * len(images_this_chunk)
+                inputs = processor(text=ners_batch,
+                                   images=images_this_chunk,
+                                   return_tensors='pt',
+                                   padding=True,
+                                   truncation=True)
+
+                with torch.no_grad():
+                    outputs = model(**inputs)
+                    target_sizes = torch.tensor(
+                        [img.size[::-1] for img in images_this_chunk])
+                    results = processor.post_process_object_detection(
+                        outputs,
+                        threshold=self.conf_thr,
+                        target_sizes=target_sizes)
+
+                image_recalls = []
+                for idx, result in enumerate(results):
+                    scores = result['scores']
+                    labels = result['labels']
+                    boxes = result['boxes']
+
+                    # sort by the confidence scores
+                    # and only keep the first num_ners predictions
+                    order_idx = scores.argsort(descending=True)
+                    scores = scores[order_idx].tolist()[:num_ners]
+                    labels = labels[order_idx].tolist()[:num_ners]
+                    boxes = boxes[order_idx].tolist()[:num_ners]
+
+                    image_area = target_sizes[idx].prod()
+                    hit = {}
+                    for box, label, score in zip(boxes, labels, scores):
+                        # this ner is already hit
+                        if ners_this_chunk[label] in hit:
+                            continue
+                        # skip boxes nearly cover the whole image
+                        xmin, ymin, xmax, ymax = box
+                        box_area = (xmax - xmin) * (ymax - ymin)
+                        if 1.0 * box_area / image_area > \
+                                self.large_area_ratio_thr:
+                            continue
+                        # skip overlapped boxes with nms-like method
+                        suppressed = False
+                        for ner in hit:
+                            if iou(box, hit[ner][0]) > self.iou_thr:
+                                suppressed = True
+                                break
+                        if suppressed:
+                            continue
+
+                        # record the new hit box
+                        hit[ners_this_chunk[label]] = (box, score)
+
+                    recall = 1.0 * len(hit) / num_ners
+                    image_recalls.append(recall)
+
+                if self.reduce_mode == 'avg':
+                    image_recall = sum(image_recalls) / len(image_recalls)
+                elif self.reduce_mode == 'max':
+                    image_recall = max(image_recalls)
+                else:
+                    image_recall = min(image_recalls)
+
+                recalls.append(image_recall)
+            offset += count
+        sample[Fields.stats][StatsKeys.phrase_grounding_recall] = recalls
+
+        return sample
+
+    def process(self, sample):
+        recalls = sample[Fields.stats][StatsKeys.phrase_grounding_recall]
+        if len(recalls) <= 0:
+            return True
+
+        keep_bools = np.array([
+            self.min_recall <= recall <= self.max_recall for recall in recalls
+        ])
+
+        # different strategies
+        if self.any:
+            return keep_bools.any()
+        else:
+            return keep_bools.all()

data_juicer/ops/filter/phrase_grounding_recall_filter.py

@@ -129,6 +129,7 @@ class StatsKeysConstant(object):
     # multimodal
     image_text_similarity = 'image_text_similarity'
     image_text_matching_score = 'image_text_matching_score'
+    phrase_grounding_recall = 'phrase_grounding_recall'
 
 
 class StatsKeys(object, metaclass=StatsKeysMeta):

data_juicer/utils/constant.py

@@ -26,30 +26,21 @@ def get_special_tokens():
 
 def remove_special_tokens(text):
     for value in get_special_tokens().values():
-        text = text.replace(value, '')
+        text = text.replace(value, '').strip()
     return text
 
 
+# Image
 def load_images(paths):
     return [load_image(path) for path in paths]
 
 
-def load_audios(paths):
-    return [load_audio(path) for path in paths]
-
-
 def load_image(path):
     img_feature = Image()
     img = img_feature.decode_example(img_feature.encode_example(path))
     return img
 
 
-def load_audio(path, sampling_rate=None):
-    aud_feature = Audio(sampling_rate)
-    aud = aud_feature.decode_example(aud_feature.encode_example(path))
-    return (aud['array'], aud['sampling_rate'])
-
-
 def pil_to_opencv(pil_image):
     if pil_image.mode != 'RGB':
         pil_image = pil_image.convert('RGB')
@@ -64,6 +55,32 @@ def get_image_size(path, ):
     return os.path.getsize(path)
 
 
+def iou(box1, box2):
+    x1_min, y1_min, x1_max, y1_max = box1
+    x2_min, y2_min, x2_max, y2_max = box2
+    area1 = (x1_max - x1_min) * (y1_max - y1_min)
+    area2 = (x2_max - x2_min) * (y2_max - y2_min)
+    ix_min = max(x1_min, x2_min)
+    ix_max = min(x1_max, x2_max)
+    iy_min = max(y1_min, y2_min)
+    iy_max = min(y1_max, y2_max)
+    intersection = max(0, (ix_max - ix_min) * (iy_max - iy_min))
+    union = area1 + area2 - intersection
+    return 1.0 * intersection / union
+
+
+# Audio
+def load_audios(paths):
+    return [load_audio(path) for path in paths]
+
+
+def load_audio(path, sampling_rate=None):
+    aud_feature = Audio(sampling_rate)
+    aud = aud_feature.decode_example(aud_feature.encode_example(path))
+    return aud['array'], aud['sampling_rate']
+
+
+# Others
 def size_to_bytes(size):
     alphabets_list = [char for char in size if char.isalpha()]
     numbers_list = [char for char in size if char.isdigit()]