Poster2plot

🚀 Poster2Plot

Poster2Plot 是一个图像描述模型，可根据电影或电视剧海报生成剧情简介。目前它生成的剧情表现尚可，但仍有提升空间。我们正在持续努力改进该模型。

🚀 快速开始

在线演示

你可以在 Hugging Face Spaces 上体验实时演示：https://huggingface.co/spaces/deepklarity/poster2plot

✨ 主要特性

该模型以 Vision Transformer (ViT) 作为图像编码器，GPT - 2 作为解码器。具体使用的模型如下：

• 编码器：google/vit-base-patch16-224-in21k
• 解码器：gpt2

📦 数据集

模型使用了公开可用的 IMDb 数据集进行训练。

💻 使用示例

基础用法

以下是在 PyTorch 中使用该模型的示例代码：

import torch
import re
import requests
from PIL import Image
from transformers import AutoTokenizer, AutoFeatureExtractor, VisionEncoderDecoderModel

# Pattern to ignore all the text after 2 or more full stops
regex_pattern = "[.]{2,}"


def post_process(text):
    try:
        text = text.strip()
        text = re.split(regex_pattern, text)[0]
    except Exception as e:
        print(e)
        pass
    return text


def predict(image, max_length=64, num_beams=4):
    pixel_values = feature_extractor(images=image, return_tensors="pt").pixel_values
    pixel_values = pixel_values.to(device)

    with torch.no_grad():
        output_ids = model.generate(
            pixel_values,
            max_length=max_length,
            num_beams=num_beams,
            return_dict_in_generate=True,
        ).sequences

    preds = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
    pred = post_process(preds[0])

    return pred


model_name_or_path = "deepklarity/poster2plot"
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# Load model.

model = VisionEncoderDecoderModel.from_pretrained(model_name_or_path)
model.to(device)
print("Loaded model")

feature_extractor = AutoFeatureExtractor.from_pretrained(model.encoder.name_or_path)
print("Loaded feature_extractor")

tokenizer = AutoTokenizer.from_pretrained(model.decoder.name_or_path, use_fast=True)
if model.decoder.name_or_path == "gpt2":
    tokenizer.pad_token = tokenizer.eos_token

print("Loaded tokenizer")

url = "https://upload.wikimedia.org/wikipedia/en/2/26/Moana_Teaser_Poster.jpg"
with Image.open(requests.get(url, stream=True).raw) as image:
    pred = predict(image)

print(pred)