Spaces:

DDingcheol
/

VaqAndOkvqa

Runtime error

App Files Files Community

VaqAndOkvqa / app.py

DDingcheol

Update app.py

b28441d 6 months ago

raw history blame contribute delete

No virus

4.21 kB

	#허깅페이스에서 돌아갈 수 있도록 바꾸어 보았음
	import torch
	from transformers import BertTokenizerFast, BertForQuestionAnswering, Trainer, TrainingArguments
	from datasets import load_dataset
	from collections import defaultdict

	# 데이터 불러오기
	dataset_load = load_dataset('Multimodal-Fatima/OK-VQA_train')
	dataset = dataset_load['train'].select(range(300))

	# 불필요한 특성 선택
	selected_features = ['image', 'answers', 'question']
	selected_dataset = dataset.map(lambda ex: {feature: ex[feature] for feature in selected_features})

	# 소프트 인코딩
	answers_to_id = defaultdict(lambda: len(answers_to_id))
	selected_dataset = selected_dataset.map(lambda ex: {
	'answers': [answers_to_id[ans] for ans in ex['answers']],
	'question': ex['question'],
	'image': ex['image']
	})

	id_to_answers = {v: k for k, v in answers_to_id.items()}
	id_to_labels = {k: ex['answers'] for k, ex in enumerate(selected_dataset)}

	selected_dataset = selected_dataset.map(lambda ex: {'answers': id_to_labels.get(ex['answers'][0]),
	'question': ex['question'],
	'image': ex['image']})

	flattened_features = []

	for ex in selected_dataset:
	flattened_example = {
	'answers': ex['answers'],
	'question': ex['question'],
	'image': ex['image'],
	}
	flattened_features.append(flattened_example)

	# 모델 가져오기
	from transformers import AutoModelForSequenceClassification, TrainingArguments, Trainer

	model_name = 'microsoft/git-base-vqav2'
	model = AutoModelForSequenceClassification.from_pretrained(model_name)

	# Trainer를 사용하여 모델 학습
	tokenizer = BertTokenizerFast.from_pretrained('bert-base-multilingual-cased')

	def preprocess_function(examples):
	tokenized_inputs = tokenizer(examples['question'], truncation=True, padding=True)
	return {
	'input_ids': tokenized_inputs['input_ids'],
	'attention_mask': tokenized_inputs['attention_mask'],
	'pixel_values': [(4, 3, 244, 244)] * len(tokenized_inputs['input_ids']),
	'pixel_mask': [1] * len(tokenized_inputs['input_ids']),
	'labels': [[label] for label in examples['answers']]
	}

	dataset = load_dataset("Multimodal-Fatima/OK-VQA_train")['train'].select(range(300))
	ok_vqa_dataset = dataset.map(preprocess_function, batched=True)
	ok_vqa_dataset.set_format(type='torch', columns=['input_ids', 'attention_mask', 'pixel_values', 'pixel_mask', 'labels'])

	training_args = TrainingArguments(
	output_dir='./results',
	num_train_epochs=20,
	per_device_train_batch_size=4,
	logging_steps=500,
	)

	trainer = Trainer(
	model=model,
	args=training_args,
	train_dataset=ok_vqa_dataset
	)

	# 모델 학습
	trainer.train()


	import gradio as gr
	import torch
	from transformers import BertTokenizer, BertForSequenceClassification

	# 모델 초기화 및 가중치 불러오기
	model_name = 'microsoft/git-base-vqav2' # 사용할 모델의 이름
	model = BertForSequenceClassification.from_pretrained(model_name)
	tokenizer = BertTokenizer.from_pretrained(model_name)

	# 예측 함수 정의
	def predict_answer(image, question):
	inputs = tokenizer(question, return_tensors='pt')
	input_ids = inputs['input_ids']
	attention_mask = inputs['attention_mask']

	# 이미지와 관련된 처리 수행
	# 이미지 처리 코드를 여기에 추가해야 합니다 (입력된 이미지에 대한 전처리 등)

	# 모델에 입력 데이터를 전달하여 예측 수행
	with torch.no_grad():
	outputs = model(input_ids=input_ids, attention_mask=attention_mask)

	# 예측 결과에서 가장 높은 확률을 가진 레이블 ID 가져오기
	predicted_label_id = torch.argmax(outputs.logits).item()
	predicted_label = id_to_label_fn(predicted_label_id)

	return predicted_label

	iface = gr.Interface(
	fn=predict_answer,
	inputs=["image", "text"],
	outputs="text",
	title="Visual Question Answering",
	description="Input an image and a question to get the model's answer.",
	example=[
	"https://your_image_url.jpg",
	"What is shown in the image?"
	]
	)

	iface.launch()