Repository files navigation

functis is a universal image reader designed for seamless reading of images into multiple formats, including PIL Images, torch Tensors via torchvision.io, and NumPy arrays via OpenCV. It offers flexibility in device placement, image format, color modes, data types, and output types.

• Read Modes: Supports multiple reading backends includingtorch,np,pil, andnvjpeg.

  note: torch is torchvision.io, np uses cv2, and nvjpeg is a private library I have yet to expose publically.

• Device Placement: Place your data on any CUDA device or CPU.
• Image Format: Choose betweenhwc orchw formats.
• Color Modes: Multiple color options includingrgb,bgr,rgba,bgra, andgrayscale.
• Data Types: Supportsuint8,float32, andfloat16 fortorch,nvjpeg andnp backends.
• Output Types: Get the result inpil,np, ortorch formats.

git clone https://github.com/aredden/functiscd functispython -m pip install.

fromfunctisimportImageReaderimreader=ImageReader().with_reader("torch").with_device("cuda:0").with_format("hwc")image=imreader.read("./your_image.jpg")

forimageinimreader.read_dir("./your_image_directory"):# process or view your image    ...# For recursive reading...forimageinimreader.read_dir("./your_image_directory",recursive=True):# process or view your image    ...# Also will optionally include paths...forpath,imageinimreader.read_dir("./your_image_directory",return_paths=True):# process or view your image    ...

Can also return paths, similar to read_dir.

forimageinimreader.read_list(list_of_your_images):# process or view your image    ...

Relatively new feature (as of oct 11 2023)

withopen("./your_image.jpg","rb")asf:image=imreader.read(f.read())

Inspiration for this idea is that I wanted to have a way to efficienctly read images of all types from a web api, here is a simple fastapi example:

fromtypingimportAnnotatedfromfastapiimportFastAPI,Bodyapp=FastAPI()inference_client= ...@app.post("/inference",tags=["inference"])asyncdefinference(image_bytes:Annotated[bytes,Body(media_type="application/octet-stream")]):globalinference_clientimage=imreader.read(image_bytes)withtorch.no_grad():results_json=inference_client.infer(image)returnresults_json

• Read Modes:torch,np,pil,nvjpeg # torch nvjpeg reader is not public yet.
• Device:cuda:<index>,cuda,cpu
• Format:hwc,chw
• Color Mode:rgb,bgr,rgba,bgra,grayscale
• Data Type:uint8,float32,float16 (only for torch and np)
• Image Type:pil,np,torch

We welcome contributions! Please see ourcontributing guidelines for more details.