Confused about dataset

[chenj02]

Great work! But I'm a litle confused about training data. It seems that StyleShot need data like, content-style-stylized image, but in the data tutorial I didn't see this preparation step? Can you explain more clearly? Thanks!

[Jeoyal]

Hi @chenj02 , thank your for your interest in our work. StyleShot follows the self-supervised training paradigm, which means reference image-style image-ground truth are processed from the same image input.
Specifically, when training stage 1, the input image will be partitioned and further processed into style embeddings by our style-aware encoder.
And in stage 2, the input image is contoured into a content image.
Although the style and content inputs are derived from the same image, they are decoupled to distinctly represent style and content, respectively.

[chenj02]

Hi @chenj02 , thank your for your interest in our work. StyleShot follows the self-supervised training paradigm, which means reference image-style image-ground truth are processed from the same image input. Specifically, when training stage 1, the input image will be partitioned and further processed into style embeddings by our style-aware encoder. And in stage 2, the input image is contoured into a content image. Although the style and content inputs are derived from the same image, they are decoupled to distinctly represent style and content, respectively.

oh! Thank you for quick reply! I understand it! I have one more question: will the StyleGallery be fully organized and open-sourced?

[Jeoyal]

Hi, we have released the JSON file and dataset tutorial of StyleGallery.
StyleGallery includes three open-sourced dataset JourneyDB, WIKIART, and a subset of stylized images from MultiGen-20M ( A subset of LAION-Aesthetics).
Before proceeding StyleGallery with further processing, you need to download these three datasets.

[chenj02]

Thanks very much!

[Delicious-Bitter-Melon]

Hi, we have released the JSON file and dataset tutorial of StyleGallery. StyleGallery includes three open-sourced dataset JourneyDB, WIKIART, and a subset of stylized images from MultiGen-20M ( A subset of LAION-Aesthetics). Before proceeding StyleGallery with further processing, you need to download these three datasets.

The URL about [WIKIART (https://huggingface.co/datasets/huggan/wikiart) is used to download the dataset in parquet format, do you provide the code to convert it to the PNG or JPG format as you use?

[Jeoyal]

Hi, we have released the JSON file and dataset tutorial of StyleGallery. StyleGallery includes three open-sourced dataset JourneyDB, WIKIART, and a subset of stylized images from MultiGen-20M ( A subset of LAION-Aesthetics). Before proceeding StyleGallery with further processing, you need to download these three datasets.

The URL about [WIKIART (https://huggingface.co/datasets/huggan/wikiart) is used to download the dataset in parquet format, do you provide the code to convert it to the PNG or JPG format as you use?

Just read the images sequentially from the .parquet file and save them.

[Delicious-Bitter-Melon]

Hi, we have released the JSON file and dataset tutorial of StyleGallery. StyleGallery includes three open-sourced dataset JourneyDB, WIKIART, and a subset of stylized images from MultiGen-20M ( A subset of LAION-Aesthetics). Before proceeding StyleGallery with further processing, you need to download these three datasets.

The URL about [WIKIART (https://huggingface.co/datasets/huggan/wikiart) is used to download the dataset in parquet format, do you provide the code to convert it to the PNG or JPG format as you use?

Just read the images sequentially from the .parquet file and save them.

Thanks for your quick reply. Is the image indexed starting from 00000.jpg or 00001.jpg?

[Jeoyal]

Hi, we have released the JSON file and dataset tutorial of StyleGallery. StyleGallery includes three open-sourced dataset JourneyDB, WIKIART, and a subset of stylized images from MultiGen-20M ( A subset of LAION-Aesthetics). Before proceeding StyleGallery with further processing, you need to download these three datasets.

The URL about [WIKIART (https://huggingface.co/datasets/huggan/wikiart) is used to download the dataset in parquet format, do you provide the code to convert it to the PNG or JPG format as you use?

Just read the images sequentially from the .parquet file and save them.

Thanks for your quick reply. Is the image indexed starting from 00000.jpg or 00001.jpg?

00000.jpg

[Jeoyal]

Hi, we have released the JSON file and dataset tutorial of StyleGallery. StyleGallery includes three open-sourced dataset JourneyDB, WIKIART, and a subset of stylized images from MultiGen-20M ( A subset of LAION-Aesthetics). Before proceeding StyleGallery with further processing, you need to download these three datasets.

The URL about [WIKIART (https://huggingface.co/datasets/huggan/wikiart) is used to download the dataset in parquet format, do you provide the code to convert it to the PNG or JPG format as you use?

Here is my load script:

class MyDataset(torch.utils.data.Dataset):

def __init__(self, path):
    super().__init__()
    self.path = path
    self.pars = os.listdir(self.path)
    self.data_paths = []
    for p in self.pars:
        table = parquet.read_table(os.path.join(self.path, p))
        df = table.to_pandas()
        for i in range(len(df)):
            self.data_paths.append([p, i])

def __getitem__(self, idx):
    par, index = self.data_paths[idx][0], self.data_paths[idx][1]
    data = parquet.read_table(os.path.join(self.path, par)).to_pandas()
    data = data.iloc[index]
    img_byte = data['image']['bytes']
    bytes_stream = BytesIO(img_byte)
    image = Image.open(bytes_stream)
    return image, data['artist'], data['genre'],  data['style']

def __len__(self):
    return len(self.data_paths)

[Delicious-Bitter-Melon]

Hi, we have released the JSON file and dataset tutorial of StyleGallery. StyleGallery includes three open-sourced dataset JourneyDB, WIKIART, and a subset of stylized images from MultiGen-20M ( A subset of LAION-Aesthetics). Before proceeding StyleGallery with further processing, you need to download these three datasets.

The URL about [WIKIART (https://huggingface.co/datasets/huggan/wikiart) is used to download the dataset in parquet format, do you provide the code to convert it to the PNG or JPG format as you use?

Here is my load script:

class MyDataset(torch.utils.data.Dataset):

def __init__(self, path):
    super().__init__()
    self.path = path
    self.pars = os.listdir(self.path)
    self.data_paths = []
    for p in self.pars:
        table = parquet.read_table(os.path.join(self.path, p))
        df = table.to_pandas()
        for i in range(len(df)):
            self.data_paths.append([p, i])

def __getitem__(self, idx):
    par, index = self.data_paths[idx][0], self.data_paths[idx][1]
    data = parquet.read_table(os.path.join(self.path, par)).to_pandas()
    data = data.iloc[index]
    img_byte = data['image']['bytes']
    bytes_stream = BytesIO(img_byte)
    image = Image.open(bytes_stream)
    return image, data['artist'], data['genre'],  data['style']

def __len__(self):
    return len(self.data_paths)

When I use the "os.listdir(self.path)", I get the sequence starting from train-00067-of-00072.parquet instead of train-00000-of-00072.parquet. So do you begin to extract images from train-00067-of-00072.parquet?

[Jeoyal]

Hi, we have released the JSON file and dataset tutorial of StyleGallery. StyleGallery includes three open-sourced dataset JourneyDB, WIKIART, and a subset of stylized images from MultiGen-20M ( A subset of LAION-Aesthetics). Before proceeding StyleGallery with further processing, you need to download these three datasets.

The URL about [WIKIART (https://huggingface.co/datasets/huggan/wikiart) is used to download the dataset in parquet format, do you provide the code to convert it to the PNG or JPG format as you use?

Here is my load script:
class MyDataset(torch.utils.data.Dataset):

def __init__(self, path):
    super().__init__()
    self.path = path
    self.pars = os.listdir(self.path)
    self.data_paths = []
    for p in self.pars:
        table = parquet.read_table(os.path.join(self.path, p))
        df = table.to_pandas()
        for i in range(len(df)):
            self.data_paths.append([p, i])

def __getitem__(self, idx):
    par, index = self.data_paths[idx][0], self.data_paths[idx][1]
    data = parquet.read_table(os.path.join(self.path, par)).to_pandas()
    data = data.iloc[index]
    img_byte = data['image']['bytes']
    bytes_stream = BytesIO(img_byte)
    image = Image.open(bytes_stream)
    return image, data['artist'], data['genre'],  data['style']

def __len__(self):
    return len(self.data_paths)

When I use the "os.listdir(self.path)", I get the sequence starting from train-00067-of-00072.parquet instead of train-00000-of-00072.parquet. So do you begin to extract images from train-00067-of-00072.parquet?

I am start from 00000 to 00072. You might need to sorted.