enczoo 0.1.2

Map images (as `PIL.Images`) to intermediate representations (as `np.ndarray`) from off-the-shelf vision models.

enczoo: a zoo of encoding models for images

enczoo is a Python library with a single goal: to map images (as PIL.Images) to features (as np.ndarray) from intermediate layers of off-the-shelf vision models, such as AlexNet and ResNet50.

This library is meant for those who need to compute off-the-shelf image features once for their project (and perhaps cache them elsewhere).

Installation

enczoo requires Python 3.12 or above, and it's recommended you use the wonderful uv to install it. Assuming you have uv, just run the following command in your project:

uv add enczoo

You can also install enczoo using pip by running:

pip install enczoo

Usage

import enczoo
import PIL.Image
image = PIL.Image.open('my-image.png')

model = enczoo.ResNet50(layer_name='avgpool') # try layer4, layer3, ...
features = model.compute_features(images=[image]) # np.ndarray

# Want another layer? Check out: print(enczoo.ResNet50.layer_names)

Why develop enczoo?

enczoo solves several tiny problems which make correctly computing image features more annoying and error-prone than it should be. For example, enczoo automatically:

• performs model-specific image normalization ("was it -1 to 1, 0 to 1, or 0-255...? ImageNet channel normalization...?"),
• correctly encodes images ("my image was in mode L, not RGB!")
• turns off batch normalization ("was the model in training mode...?")
• extracts intermediate layers by name ("how do I do that forward hook thing again...?")
• turns off autograd, and returns tensors as np.ndarray (no more .cpu().numpy())
• performs image cropping to fit images to the expected input tensor shape
• and more!