基于GPT-2实现图像文本生成

原理

使用GPT-2模型处理文本，做decoder。
使用google的vit-base-patch16-224模型处理图像，做encoder。
最后通过VisionEncoderDecoderModel将这两个模型粘起来。
如下图所示。

代码

1、导入相关的包

from transformers import (VisionEncoderDecoderModel, 
                          AutoTokenizer,ViTImageProcessor)
import torch
from PIL import Image

2、加载模型

# VIT_MODEL_NAME_OR_PATH = "./vit-base-patch16-224"
# GPT_MODEL_NAME_OR_PATH = "./gpt2_chinese"
# vision_encoder_decoder_model_name_or_path = "./vit-gpt2-image-chinese-captioning"

VIT_MODEL_NAME_OR_PATH = "google/vit-base-patch16-224"
GPT_MODEL_NAME_OR_PATH = "yuanzhoulvpi/gpt2_chinese"
vision_encoder_decoder_model_name_or_path = "vit-gpt2-image-chinese-captioning"

processor = ViTImageProcessor.from_pretrained(VIT_MODEL_NAME_OR_PATH)
tokenizer = AutoTokenizer.from_pretrained(GPT_MODEL_NAME_OR_PATH)
model = VisionEncoderDecoderModel.from_pretrained(vision_encoder_decoder_model_name_or_path)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

如果网速过慢，可以先把模型下载到本地，下载方法：下载huggingface-transformers模型至本地，并使用from_pretrained方法加载

3、推理，生成文本

max_length = 16
num_beams = 4
gen_kwargs = {"max_length": max_length, "num_beams": num_beams}


def predict_step(image_paths):
    images = []
    for image_path in image_paths:
        i_image = Image.open(image_path)
        if i_image.mode != "RGB":
            i_image = i_image.convert(mode="RGB")

        images.append(i_image)

    pixel_values = processor(images=images, return_tensors="pt").pixel_values
    pixel_values = pixel_values.to(device)

    output_ids = model.generate(pixel_values, **gen_kwargs)

    preds = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
    preds = [pred.strip() for pred in preds]
    return preds


predict_step(['./images/a.jpg'])

4、效果
我上传的图片为：
生成的文本为：

参考github开源项目