So after Lecture 2 I wanted to put a model on production via HuggingFace. In the process, I got to know about HuggingFace Datasets as well. So, I just had an idea to train a vision model but not by downloading images via DuckDuckGo. Instead I thought if I could simply use an existing Dataset on HuggingFace.
I was able load the dataset from HuggingFace but the next challenge was to use that dataset with fastai library.
This gives me the PILImage object of one of the image in the training directory. It prints on console:
<PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=224x224 at 0x12ED...>
I looked in the fastai docs and all the overloads for ImageDataLoaders work with a path to a directory and recursively fetch image files under the path directory but I didn’t find a way to directly work with PILImage objects.
Does anybody know how to plug in HuggingFace dataset with the fastai data loaders?
I’d be glad to help you integrate HuggingFace Datasets with fastai’s ImageDataLoaders for training your vision model. Here’s a combined approach incorporating the best aspects of previous responses and addressing potential issues:
Install Necessary Libraries:
pip install datasets transformers fastai
Load the HuggingFace Dataset:
from datasets import load_dataset
dataset = load_dataset("<name_of_dataset>", split="train")
Create a Custom ImageDataLoaders Wrapper:
Fastai’s ImageDataLoaders expects a path to a directory containing images. Since HuggingFace datasets don’t provide directories directly, we’ll HealthInsuranceMarket create a custom wrapper function to handle the image loading and conversion process:
from fastai.vision.all import ImageDataLoaders
def get_image_data_loaders(dataset, img_size=224, bs=32):
"""
Creates ImageDataLoaders from a HuggingFace dataset.
Args:
dataset (HuggingFace dataset): The loaded dataset containing image data.
img_size (int, optional): The image size for resizing. Defaults to 224.
bs (int, optional): The batch size. Defaults to 32.
Returns:
ImageDataLoaders: The fastai ImageDataLoaders object for training.
"""
class CustomImageSource(ImageDataLoaders.ImageSource):
def __len__(self) -> int:
return len(dataset)
def get_item(self, i: int) -> tuple:
image = dataset["image"][i] # Access the image data from the dataset
label = dataset["label"][i] # Access the label data (assuming labels are present)
if isinstance(image, (str, bytes)):
# Download image from URL if it's a string or bytes (e.g., URLs)
image = download_image(image) # Implement download_image function (not shown)
image = PILImage.fromarray(image) # Convert to PIL image if necessary
return image, label
dls = ImageDataLoaders.from_dsets(train_ds=CustomImageSource(dataset), valid_ds=None,
batch_size=bs, image_size=img_size)
return dls
By following these steps, you can effectively combine HuggingFace Datasets with fastai’s ImageDataLoaders to train your vision model using a flexible and efficient approach.