nwf-vision 0.1.0

NWF for computer vision - ConvVAE, pretrained encoders, continual learning

nwf-vision

NWF for computer vision: ConvVAE, pretrained encoders, continual learning on images.

Installation

pip install nwf-vision
# Requires: nwf-core, torch, torchvision

Quick start

from nwf.vision import ConvVAEEncoder
from nwf import Charge, Field
import numpy as np

enc = ConvVAEEncoder(input_shape=(3, 32, 32), latent_dim=64)
enc.fit(images, epochs=10)

z, sigma = enc.encode(images[:5])
field = Field()
for i in range(5):
    field.add(Charge(z=z[i], sigma=sigma[i]), labels=[labels[i]])

Components

• ConvVAEEncoder - convolutional VAE for images (CIFAR, MNIST)
• PretrainedVisionEncoder - ResNet/EfficientNet + (z, sigma) head
• Examples: Split-CIFAR-10, OOD detection, active learning

Run example

pip install nwf-core nwf-vision
python examples/split_cifar.py --epochs 5

Note: if you have nwf-research in PYTHONPATH, it may shadow nwf-core. Use a clean env or install nwf-core in development mode.

License

MIT