pandas-image-methods 0.2.0

Image methods for pandas dataframes using Pillow

Pandas Image Methods

Image methods for pandas dataframes using Pillow.

Features:

• Use PIL.Image objects in pandas dataframes
• Call PIL.Image methods on a column, for example:
  • .crop()
  • .filter()
  • .resize()
  • .rotate()
  • .transpose()
• Save dataframes with PIL.Image objects to Parquet
• Process images in parallel with Dask

Installation

pip install pandas-image-methods

Usage

You can open images as PIL.Image objects using the .open() method.

Once the images are opened, you can call any PIL Image method:

import pandas as pd
from pandas_image_methods import PILMethods

pd.api.extensions.register_series_accessor("pil")(PILMethods)

df = pd.DataFrame({"file_path": ["path/to/image.png"]})
df["image"] = df["file_path"].pil.open()
df["image"] = df["image"].pil.rotate(90)
# 0    <PIL.Image.Image size=200x200>
# Name: image, dtype: object, PIL methods enabled

Here is how to enable PIL methods for PIL Images created manually:

df = pd.DataFrame({"image": [PIL.Image.open("path/to/image.png")]})
df["image"] = df["image"].pil.enable()
df["image"] = df["image"].pil.rotate(90)
# 0    <PIL.Image.Image size=200x200>
# Name: image, dtype: object, PIL methods enabled

Save

You can save a dataset of PIL Images to Parquet:

# Save
df = pd.DataFrame({"file_path": ["path/to/image.png"]})
df["image"] = df["file_path"].pil.open()
df.to_parquet("data.parquet")

# Later
df = pd.read_parquet("data.parquet")
df["image"] = df["image"].pil.enable()

This doesn't just save the paths to the image files, but the actual images themselves !

Under the hood it saves dictionaries of {"bytes": <bytes of the image file>, "path": <path or name of the image file>}. The images are saved as bytes using their image encoding or PNG by default. Anyone can load the Parquet data even without pandas-image-methods since it doesn't rely on extension types.

Note: if you created the PIL Images manually, don't forget to enable the PIL methods to enable saving to Parquet.

Run in parallel

Dask DataFrame parallelizes pandas to handle large datasets. It enables faster local processing with multiprocessing as well as distributed large scale processing. Dask mimics the pandas API:

import dask.dataframe as dd
from pandas_image_methods import PILMethods

dd.api.extensions.register_series_accessor("pil")(PILMethods)

df = dd.read_csv("path/to/large/dataset.csv")
df = df.repartition(npartitions=1000)  # divide the processing in 1000 jobs

df["image"] = df["file_path"].pil.open()
df["image"] = df["image"].pil.rotate(90)
df["image"].head(1)
# 0    <PIL.Image.Image size=200x200>
# Name: image, dtype: object, PIL methods enabled

Hugging Face support

Most image datasets in Parquet format on Hugging Face are compatible with pandas-image-methods. For example you can load the CIFAR-100 dataset:

df = pd.read_parquet("hf://datasets/uoft-cs/cifar100/cifar100/train-00000-of-00001.parquet")
df["image"] = df["image"].pil.open()

Datasets created with pandas-image-methods and saved to Parquet are compatible with the Dataset Viewer on Hugging Face and the datasets library.

Display in Notebooks

You can display a pandas dataframe of images in a Jupyter Notebook or on Google Colab in HTML:

from IPython.display import HTML
HTML(df.head().to_html(escape=False, formatters={"image": df.image.pil.html_formatter}))

Example on the julien-c/impressionists dataset for painting classification: