nwf-vision 0.1.2

NWF for computer vision: ConvVAE, ResNet encoders, Split-CIFAR incremental learning, continual learning on images.

nwf-vision

NWF for Computer Vision

nwf-vision provides encoders and examples for incremental learning, OOD detection, and active learning on images. Built on top of nwf-core, it uses its indices, metrics, and calibrators.

Features

• ConvVAEEncoder — convolutional VAE for images (CIFAR 32x32, MNIST)
• PretrainedVisionEncoder — ResNet18/34 with (z, sigma) head for NWF charges
• Split-CIFAR-10 example — incremental classification without catastrophic forgetting
• OOD detection example — CIFAR-10 (in) vs SVHN (out) using semantic potential
• Active learning example — uncertainty sampling vs random on CIFAR-10
• Support for NHWC and NCHW input layouts

---

Installation

pip install nwf-vision

Requires: nwf-core, torch, torchvision.

---

Encoders

ConvVAEEncoder

Convolutional VAE producing (z, sigma) for NWF charges. Suitable for small images (32x32).

from nwf.vision import ConvVAEEncoder
import numpy as np

enc = ConvVAEEncoder(input_shape=(3, 32, 32), latent_dim=64)
enc.fit(train_images, epochs=20)
z, sigma = enc.encode(images[:10])

• input_shape — (C, H, W), e.g. (3, 32, 32) for CIFAR
• latent_dim — dimension of z
• hidden_dims — encoder channel sizes (default: 32, 64, 128)
• fit(train_data, epochs, batch_size, lr) — train VAE
• encode(x) — returns (z, sigma) as numpy arrays; accepts NHWC or NCHW

PretrainedVisionEncoder

ResNet backbone from torchvision with a head producing (z, sigma).

from nwf.vision import PretrainedVisionEncoder

enc = PretrainedVisionEncoder(
    backbone="resnet18",
    latent_dim=64,
    pretrained=True,
    trainable=False,
)
enc.fit(images, epochs=5)
z, sigma = enc.encode(images)

• backbone — "resnet18" or "resnet34"
• pretrained — use ImageNet weights
• trainable — fine-tune backbone or freeze it
• Input: (N, 3, 224, 224) for ImageNet-style models

---

Example: Split-CIFAR-10

Incremental learning: 5 tasks, 2 classes each. Train ConvVAE once, add charges per task.

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

The example demonstrates adding new classes without retraining the encoder and without catastrophic forgetting.

---

Examples: OOD and Active Learning

python examples/ood_detection.py      # CIFAR-10 vs SVHN, AUROC
python examples/active_learning.py    # Uncertainty vs random sampling

---

Links

• Article (Habr): Нейровесовые Поля (NWF) — теория, continual learning
• Core: nwf-core
• Research: nwf-research

License

MIT

---

nwf-vision (Русский)

NWF для компьютерного зрения

nwf-vision предоставляет энкодеры и примеры для инкрементального обучения, OOD-детекции и активного обучения на изображениях.

Компоненты

• ConvVAEEncoder — свёрточный VAE для изображений (CIFAR 32x32, MNIST); выход (z, sigma)
• PretrainedVisionEncoder — ResNet18/34 с головой для (z, sigma); предобученные веса
• Split-CIFAR-10 — пример инкрементальной классификации без катастрофического забывания
• OOD detection — пример CIFAR-10 vs SVHN через потенциал
• Active learning — uncertainty sampling vs random

Установка

pip install nwf-vision

Пример

from nwf.vision import ConvVAEEncoder
from nwf import Charge, Field

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]])

Лицензия

MIT