enczoo 0.1.6.dev1

Easily map images (as `PIL.Images`) to features (as `np.ndarray`) from pretrained vision models.

enczoo: easily extract image features from pretrained vision models

enczoo is a Python library with a simple goal: to make it as easy as possible to map images (as PIL.Images) to features (as numpy arrays) from state-of-the-art vision models, such as Imagenet-pretrained ResNet50 and CLIP ViT-B/16.

Installation

enczoo requires Python 3.12 or above, and is installed using the wonderful uv project manager. Once you have uv installed, just run the following command in your project:

uv add enczoo

Usage

import enczoo
from PIL import Image

image = Image.open('my-image.png')
model = enczoo.ResNet50(
    layer_name='avgpool',
    # device=gpu
) 
features = model.compute_features(images=[image]) # np.ndarray
# Want another layer? Check out: print(enczoo.ResNet50.layer_names)

Available models

Pixels

• Family: raw pixels
• Returns: float32 RGB pixels after preprocessing
• Output shape: [B, 224, 224, 3]
• Academic reference: none; this is an enczoo convenience encoder

AlexNet

• Family: ImageNet-pretrained CNN
• Returns: intermediate activations from the requested layer
• Output shape: depends on layer_name
• Layer selection: inspect enczoo.AlexNet.layer_names
• Academic reference: AlexNet, "ImageNet Classification with Deep Convolutional Neural Networks" (Krizhevsky et al., 2012)

ResNet50

• Family: ImageNet-pretrained CNN
• Returns: intermediate activations from the requested layer
• Output shape: depends on layer_name
• Layer selection: inspect enczoo.ResNet50.layer_names
• Academic reference: ResNet, "Deep Residual Learning for Image Recognition" (He et al., 2015)

RobustResNet50

• Family: adversarially robust ImageNet ResNet-50
• Returns: intermediate activations from the requested layer
• Output shape: depends on layer_name
• Layer selection: inspect enczoo.RobustResNet50.layer_names
• Weights: downloaded on first use from the released ImageNet L2 epsilon-3.0 checkpoint
• Academic reference: Engstrom et al., "Robustness (Python Library)" release checkpoint via the MadryLab model weights

ConvNeXtB

• Family: ImageNet-pretrained CNN
• Returns: intermediate activations from the requested layer
• Output shape: depends on layer_name
• Layer selection: inspect enczoo.ConvNeXtB.layer_names
• Academic reference: ConvNeXt, "A ConvNet for the 2020s" (Liu et al., 2022)

CLIPResNet50

• Family: CLIP ResNet visual encoder
• Returns: intermediate activations from the requested visual layer
• Output shape: depends on layer_name
• Layer selection: inspect enczoo.CLIPResNet50.layer_names
• Academic reference: CLIP, "Learning Transferable Visual Models From Natural Language Supervision" (Radford et al., 2021)

CLIPViTB16

• Family: CLIP vision transformer
• Returns: the model's pooled CLS-based image embedding
• Output shape: [B, 768]
• Academic reference: CLIP, "Learning Transferable Visual Models From Natural Language Supervision" (Radford et al., 2021)

DINOv2ViTB14

• Family: self-supervised vision transformer
• Returns: the model's pooled CLS-based image embedding
• Output shape: [B, 768]
• Academic reference: DINOv2, "DINOv2: Learning Robust Visual Features without Supervision" (Oquab et al., 2023)

AligNetViTB16

• Family: AlignNet-aligned vision transformer
• Returns: the SavedModel feature tensor selected from the exported pre_logits output
• Output shape: depends on the downloaded model
• Weights: downloaded on first use and cached under ENCZOO_CACHE_DIR or the platform cache directory
• Academic reference: Muttenthaler et al. 2025; weights come from the AlignNet model release

ImageNetViTB16

• Family: ImageNet-pretrained vision transformer from the AlignNet release
• Returns: the SavedModel feature tensor selected from the exported pre_logits output
• Output shape: depends on the downloaded model
• Weights: downloaded on first use and cached under ENCZOO_CACHE_DIR or the platform cache directory
• Academic reference: Muttenthaler et al. 2025; weights come from the AlignNet model release

UnaligNetViTB16

• Family: unaligned vision transformer from the AlignNet release
• Returns: the SavedModel feature tensor selected from the exported pre_logits output
• Output shape: depends on the downloaded model
• Weights: downloaded on first use and cached under ENCZOO_CACHE_DIR or the platform cache directory
• Academic reference: Muttenthaler et al. 2025; weights come from the AlignNet model release

Why develop enczoo?

Under the hood, 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 transforms ("was it -1 to 1, 0 to 1, or 0-255...?"),
• ensures images are in RGB format
• puts the model in inference, not training, mode
• turns off autograd
• returns tensors as np.ndarray (no more detach().cpu().numpy())
• resizes the image while preserving aspect ratio
• and more!