def load_item(self, idx):
sample_info = self.annotation_db[idx]
current_sample = Sample()
processed_sentence = self.text_processor(
{"text": sample_info["sentence"]})
current_sample.text = processed_sentence["text"]
if "input_ids" in processed_sentence:
current_sample.update(processed_sentence)
if self._use_features is True:
# Remove sentence id from end
identifier = "-".join(sample_info["identifier"].split("-")[:-1])
# Load img0 and img1 features
sample_info["feature_path"] = "{}-img0.npy".format(identifier)
features = self.features_db[idx]
if hasattr(self, "transformer_bbox_processor"):
features["image_info_0"] = self.transformer_bbox_processor(
features["image_info_0"])
current_sample.img0 = Sample()
current_sample.img0.update(features)
sample_info["feature_path"] = "{}-img1.npy".format(identifier)
features = self.features_db[idx]
if hasattr(self, "transformer_bbox_processor"):
features["image_info_0"] = self.transformer_bbox_processor(
features["image_info_0"])
current_sample.img1 = Sample()
current_sample.img1.update(features)
is_correct = 1 if sample_info["label"] == "True" else 0
current_sample.targets = torch.tensor(is_correct, dtype=torch.long)
return current_sample
def __getitem__(self, idx):
sample_info = self.annotation_db[idx]
current_sample = Sample()
if self._use_features:
features = self.features_db[idx]
if hasattr(self, "transformer_bbox_processor"):
features["image_info_0"] = self.transformer_bbox_processor(
features["image_info_0"])
if self.config.get("use_image_feature_masks", False):
current_sample.update({
"image_labels":
self.masked_region_processor(features["image_feature_0"])
})
current_sample.update(features)
else:
image_path = str(sample_info["image_name"]) + ".jpg"
current_sample.image = self.image_db.from_path(
image_path)["images"][0]
current_sample = self._add_masked_question(sample_info, current_sample)
if self._add_answer:
current_sample = self.add_answer_info(sample_info, current_sample)
return current_sample
def __getitem__(self, idx: int) -> Sample:
sample_info = self.annotation_db[idx]
current_sample = Sample()
processed_caption = self.masked_token_processor({
"text_a":
sample_info["caption"],
"text_b":
"",
"is_correct":
True
})
current_sample.update(processed_caption)
current_sample.image_id = sample_info["image_id"]
current_sample.feature_path = sample_info["feature_path"]
# Get the image features
if self._use_features:
features = self.features_db[idx]
image_info_0 = features["image_info_0"]
if image_info_0 and "image_id" in image_info_0.keys():
image_info_0["feature_path"] = image_info_0["image_id"]
image_info_0.pop("image_id")
current_sample.update(features)
return current_sample
def __getitem__(self, idx):
sample_info = self.annotation_db[idx]
sample_info = self.preprocess_sample_info(sample_info)
current_sample = Sample()
# breaking change from VQA2Dataset: load question_id
current_sample.question_id = torch.tensor(
sample_info["question_id"], dtype=torch.int
)
if isinstance(sample_info["image_id"], int):
current_sample.image_id = str(sample_info["image_id"])
else:
current_sample.image_id = sample_info["image_id"]
if self._use_features is True:
features = self.features_db[idx]
current_sample.update(features)
current_sample = self.add_sample_details(sample_info, current_sample)
current_sample = self.add_answer_info(sample_info, current_sample)
# only the 'max_features' key is needed
# pop other keys to minimize data loading overhead
if hasattr(current_sample, "image_info_0"):
for k in list(current_sample.image_info_0):
if k != "max_features":
current_sample.image_info_0.pop(k)
if hasattr(current_sample, "image_info_1"):
for k in list(current_sample.image_info_1):
if k != "max_features":
current_sample.image_info_1.pop(k)
return current_sample
def __getitem__(self, idx):
sample_info = self.annotation_db[idx]
sample_info = self._preprocess_answer(sample_info)
sample_info["question_id"] = sample_info["id"]
current_sample = Sample()
if self._use_features is True:
features = self.features_db[idx]
if hasattr(self, "transformer_bbox_processor"):
features["image_info_0"] = self.transformer_bbox_processor(
features["image_info_0"])
if self.config.get("use_image_feature_masks", False):
current_sample.update({
"image_labels":
self.masked_region_processor(features["image_feature_0"])
})
current_sample.update(features)
current_sample = self._add_masked_question(sample_info, current_sample)
if self._add_answer:
current_sample = self.add_answer_info(sample_info, current_sample)
return current_sample
def load_item(self, idx):
sample_info = self.annotation_db[idx]
sample_info = self.preprocess_sample_info(sample_info)
current_sample = Sample()
if self._dataset_type != "test":
text_processor_argument = {"tokens": sample_info["caption_tokens"]}
processed_caption = self.text_processor(text_processor_argument)
current_sample.text = processed_caption["text"]
current_sample.caption_id = torch.tensor(sample_info["caption_id"],
dtype=torch.int)
current_sample.caption_len = torch.tensor(len(
sample_info["caption_tokens"]),
dtype=torch.int)
current_sample.image_id = object_to_byte_tensor(
sample_info["image_id"])
if self._use_features:
features = self.features_db[idx]
current_sample.update(features)
else:
image_path = str(sample_info["image_name"]) + ".jpg"
current_sample.image = self.image_db.from_path(
image_path)["images"][0]
# Add reference captions to sample
current_sample = self.add_reference_caption(sample_info,
current_sample)
return current_sample
def load_item(self, idx):
sample_info = self.annotation_db[idx]
current_sample = Sample()
processed_sentence = self.text_processor(
{"text": sample_info["sentence2"]})
current_sample.text = processed_sentence["text"]
if "input_ids" in processed_sentence:
current_sample.update(processed_sentence)
if self._use_features is True:
# Remove sentence id from end
identifier = sample_info["Flikr30kID"].split(".")[0]
# Load img0 and img1 features
sample_info["feature_path"] = "{}.npy".format(identifier)
features = self.features_db[idx]
if hasattr(self, "transformer_bbox_processor"):
features["image_info_0"] = self.transformer_bbox_processor(
features["image_info_0"])
current_sample.update(features)
label = LABEL_TO_INT_MAPPING[sample_info["gold_label"]]
current_sample.targets = torch.tensor(label, dtype=torch.long)
return current_sample
def __getitem__(self, idx):
sample_info = self.annotation_db[idx]
current_sample = Sample()
text_processor_argument = {"text": sample_info["question_str"]}
processed_question = self.text_processor(text_processor_argument)
current_sample.text = processed_question["text"]
if "input_ids" in processed_question:
current_sample.update(processed_question)
current_sample.question_id = torch.tensor(sample_info["question_id"],
dtype=torch.int)
if isinstance(sample_info["image_id"], int):
current_sample.image_id = torch.tensor(sample_info["image_id"],
dtype=torch.int)
else:
current_sample.image_id = sample_info["image_id"]
if self._use_features is True:
features = self.features_db[idx]
if hasattr(self, "transformer_bbox_processor"):
features["image_info_0"] = self.transformer_bbox_processor(
features["image_info_0"])
current_sample.update(features)
# Depending on whether we are using soft copy this can add
# dynamic answer space
current_sample = self.add_answer_info(sample_info, current_sample)
return current_sample
def __getitem__(self, idx):
sample_info = self.annotation_db[idx]
sample_info = self.preprocess_sample_info(sample_info)
current_sample = Sample()
processed_text = self.text_processor({"text": sample_info["text"]})
current_sample.text = processed_text["text"]
if "input_ids" in processed_text:
current_sample.update(processed_text)
current_sample.id = torch.tensor(int(sample_info["id"]),
dtype=torch.int)
# Instead of using idx directly here, use sample_info to fetch
# the features as feature_path has been dynamically added
features = self.features_db.get(sample_info)
if hasattr(self, "transformer_bbox_processor"):
features["image_info_0"] = self.transformer_bbox_processor(
features["image_info_0"])
current_sample.update(features)
if "label" in sample_info:
current_sample.targets = torch.tensor(sample_info["label"],
dtype=torch.long)
return current_sample
def load_item(self, idx):
sample_info = self.annotation_db[idx]
current_sample = Sample()
if "question_tokens" in sample_info:
text_processor_argument = {
"tokens": sample_info["question_tokens"],
"text": sample_info["question_str"],
}
else:
text_processor_argument = {"text": sample_info["question"]}
processed_question = self.text_processor(text_processor_argument)
current_sample.text = processed_question["text"]
if "input_ids" in processed_question:
current_sample.update(processed_question)
current_sample.question_id = torch.tensor(
sample_info["question_id"], dtype=torch.int
)
if isinstance(sample_info["image_id"], int):
current_sample.image_id = torch.tensor(
sample_info["image_id"], dtype=torch.int
)
else:
current_sample.image_id = sample_info["image_id"]
if "question_tokens" in sample_info:
current_sample.text_len = torch.tensor(
len(sample_info["question_tokens"]), dtype=torch.int
)
if self._use_features:
features = self.features_db[idx]
if hasattr(self, "transformer_bbox_processor"):
features["image_info_0"] = self.transformer_bbox_processor(
features["image_info_0"]
)
current_sample.update(features)
else:
image_path = sample_info["image_name"] + ".jpg"
current_sample.image = self.image_db.from_path(image_path)["images"][0]
# Add details for OCR like OCR bbox, vectors, tokens here
current_sample = self.add_ocr_details(sample_info, current_sample)
# Depending on whether we are using soft copy this can add
# dynamic answer space
current_sample = self.add_answer_info(sample_info, current_sample)
return current_sample
def test_sample_working(self):
initial = Sample()
initial.x = 1
initial["y"] = 2
# Assert setter and getter
self.assertEqual(initial.x, 1)
self.assertEqual(initial["x"], 1)
self.assertEqual(initial.y, 2)
self.assertEqual(initial["y"], 2)
update_dict = {"a": 3, "b": {"c": 4}}
initial.update(update_dict)
self.assertEqual(initial.a, 3)
self.assertEqual(initial["a"], 3)
self.assertEqual(initial.b.c, 4)
self.assertEqual(initial["b"].c, 4)
def classify(self, image: ImageType, text: str):
"""Classifies a given image and text in it into Hateful/Non-Hateful.
Image can be a url or a local path or you can directly pass a PIL.Image.Image
object. Text needs to be a sentence containing all text in the image.
>>> from VisualBERT.mmf.models.mmbt import MMBT
>>> model = MMBT.from_pretrained("mmbt.hateful_memes.images")
>>> model.classify("some_url", "some_text")
{"label": 0, "confidence": 0.56}
Args:
image (ImageType): Image to be classified
text (str): Text in the image
Returns:
bool: Whether image is hateful (1) or non hateful (0)
"""
if isinstance(image, str):
if image.startswith("http"):
temp_file = tempfile.NamedTemporaryFile()
download(image,
*os.path.split(temp_file.name),
disable_tqdm=True)
image = tv_helpers.default_loader(temp_file.name)
temp_file.close()
else:
image = tv_helpers.default_loader(image)
text = self.processor_dict["text_processor"]({"text": text})
image = self.processor_dict["image_processor"](image)
sample = Sample()
sample.text = text["text"]
if "input_ids" in text:
sample.update(text)
sample.image = image
sample_list = SampleList([sample])
device = next(self.model.parameters()).device
sample_list = sample_list.to(device)
output = self.model(sample_list)
scores = nn.functional.softmax(output["scores"], dim=1)
confidence, label = torch.max(scores, dim=1)
return {"label": label.item(), "confidence": confidence.item()}
def __getitem__(self, idx: int) -> Type[Sample]:
sample_info = self.annotation_db[idx]
current_sample = Sample()
processed_question = self.text_processor(
{"text": sample_info["question"]})
current_sample.update(processed_question)
current_sample.id = torch.tensor(int(sample_info["question_id"]),
dtype=torch.int)
image_path = self.get_image_path(sample_info["image_id"],
sample_info["coco_split"])
current_sample.image = self.image_db.from_path(image_path)["images"][0]
if "answers" in sample_info:
answers = self.answer_processor(
{"answers": sample_info["answers"]})
current_sample.targets = answers["answers_scores"]
return current_sample
def __call__(self, item):
texts = item["text"]
if not isinstance(texts, list):
texts = [texts]
processed = []
for idx, text in enumerate(texts):
sample = Sample()
processed_text = self.tokenizer({"text": text})
sample.update(processed_text)
sample.segment_ids.fill_(idx)
processed.append(sample)
# Use SampleList to convert list of tensors to stacked tensors
processed = SampleList(processed)
if self.fusion_strategy == "concat":
processed.input_ids = processed.input_ids.view(-1)
processed.input_mask = processed.input_mask.view(-1)
processed.segment_ids = processed.segment_ids.view(-1)
processed.lm_label_ids = processed.lm_label_ids.view(-1)
return processed.to_dict()
def __getitem__(self, idx):
sample_info = self.annotation_db[idx]
current_sample = Sample()
plot = sample_info["plot"]
if isinstance(plot, list):
plot = plot[0]
processed_sentence = self.text_processor({"text": plot})
current_sample.text = processed_sentence["text"]
if "input_ids" in processed_sentence:
current_sample.update(processed_sentence)
if self._use_images is True:
current_sample.image = self.image_db[idx]["images"][0]
processed = self.answer_processor({"answers": sample_info["genres"]})
current_sample.answers = processed["answers"]
current_sample.targets = processed["answers_scores"]
return current_sample
def __getitem__(self, idx):
sample_info = self.annotation_db[idx]
current_sample = Sample()
processed_text = self.text_processor({"text": sample_info["text"]})
current_sample.text = processed_text["text"]
if "input_ids" in processed_text:
current_sample.update(processed_text)
current_sample.id = torch.tensor(int(sample_info["id"]),
dtype=torch.int)
# Get the first image from the set of images returned from the image_db
current_sample.image = self.image_db[idx]["images"][0]
if "label" in sample_info:
current_sample.targets = torch.tensor(sample_info["label"],
dtype=torch.long)
return current_sample
def __getitem__(self, idx):
sample_info = self.annotation_db[idx]
current_sample = Sample()
plot = sample_info["plot"]
if isinstance(plot, list):
plot = plot[0]
processed_sentence = self.text_processor({"text": plot})
current_sample.text = processed_sentence["text"]
if "input_ids" in processed_sentence:
current_sample.update(processed_sentence)
if self._use_features is True:
features = self.features_db[idx]
if hasattr(self, "transformer_bbox_processor"):
features["image_info_0"] = self.transformer_bbox_processor(
features["image_info_0"])
current_sample.update(features)
processed = self.answer_processor({"answers": sample_info["genres"]})
current_sample.answers = processed["answers"]
current_sample.targets = processed["answers_scores"]
return current_sample
def load_item(self, idx):
sample_info = self.annotation_db[idx]
current_sample = Sample()
processed_caption = self.text_processor(
{"text": sample_info["captions"][0]})
current_sample.text = processed_caption["text"]
current_sample.caption_len = torch.tensor(len(
processed_caption["text"]),
dtype=torch.int)
if isinstance(sample_info["image_id"], int):
current_sample.image_id = torch.tensor(sample_info["image_id"],
dtype=torch.int)
else:
current_sample.image_id = sample_info["image_id"]
if self._use_features is True:
features = self.features_db[idx]
current_sample.update(features)
current_sample.answers = torch.stack([processed_caption["text"]])
return current_sample