bforbuntyai 0.2.2

Run GenAI experiments in a few lines instead of hundreds

bforbuntyai

Run GenAI experiments in a few lines instead of hundreds.

from bforbuntyai import GAN, dataset
gan = GAN(dataset.FashionMNIST())
gan.train(epochs=50)
gan.generate(n=25)

Installation

# Core (no heavy ML libs)
pip install bforbuntyai

# With PyTorch (GAN, DCGAN, cGAN, VAE, Pix2Pix)
pip install "bforbuntyai[torch]"

# With TensorFlow (AutoEncoder)
pip install "bforbuntyai[tensorflow]"

# With HuggingFace (TextGenerator, TextFineTuner, ImageCaptioner, EthicalEvaluator)
pip install "bforbuntyai[transformers]"

# With Stable Diffusion
pip install "bforbuntyai[diffusers]"

# With ethical evaluation
pip install "bforbuntyai[ethics]"

# Everything at once
pip install "bforbuntyai[all]"

HuggingFace Authentication

Some models (Stable Diffusion gated versions, private datasets) require a HuggingFace token.

from bforbuntyai import auth

auth.login()                    # interactive browser prompt
auth.login(token="hf_...")      # explicit token
auth.whoami()                   # check who is logged in
auth.logout()

# Or pass token directly to any model
sd = StableDiffusion(token="hf_...")
data = dataset.HuggingFace("private/dataset", token="hf_...")

Token is also auto-read from the HF_TOKEN or HUGGINGFACE_HUB_TOKEN environment variable.

Models & Datasets

AutoEncoder (TensorFlow)

from bforbuntyai import AutoEncoder, dataset

ae = AutoEncoder(dataset.MNIST(), encoding_dim=64)
ae.train(epochs=50)
ae.visualize()              # original vs reconstructed grid
ae.save("ae.keras")

GAN — Vanilla (PyTorch)

from bforbuntyai import GAN, dataset

gan = GAN(dataset.FashionMNIST(), latent_dim=100)
gan.train(epochs=50)
gan.generate(n=25)
gan.save("gan.pth")

DCGAN — Convolutional GAN (PyTorch)

from bforbuntyai import DCGAN, dataset

dcgan = DCGAN(dataset.MNIST())
dcgan.train(epochs=20)
dcgan.generate(n=25)

Conditional GAN (PyTorch)

from bforbuntyai import ConditionalGAN, dataset

cgan = ConditionalGAN(dataset.MNIST(), num_classes=10)
cgan.train(epochs=50)
cgan.generate_class(labels=[0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

VAE — Variational AutoEncoder (PyTorch)

from bforbuntyai import VAE, dataset

vae = VAE(dataset.MNIST(), latent_dim=2)
vae.train(epochs=20)
vae.visualize()              # reconstructions
vae.visualize_latent()       # 2-D scatter plot of latent space
vae.interpolate(img_a, img_b, steps=10)    # morph between images

Pix2Pix Image Translation (PyTorch)

from bforbuntyai import Pix2Pix, dataset

p2p = Pix2Pix(dataset.Edges2Shoes())      # auto-downloads ~1.5 GB
p2p.train(epochs=5)
p2p.visualize()              # input edge | generated photo | real photo

GPT-2 Text Generation

from bforbuntyai import TextGenerator

gen = TextGenerator(model_name="gpt2")
gen.generate("Once upon a time", max_tokens=100, num_return=3)

GPT-2 Fine-tuning

from bforbuntyai import TextFineTuner, dataset

data = dataset.HuggingFace("amazon_polarity", split="train[:200]")
tuner = TextFineTuner(dataset=data)
tuner.train(epochs=3)
tuner.generate("This product is")
tuner.save("./my-gpt2")

Stable Diffusion

from bforbuntyai import StableDiffusion, auth

auth.login()    # required for some model versions
sd = StableDiffusion()
sd.generate("A sunset over mountains, 4K photo", n=4, steps=20)

BLIP Image Captioning

from bforbuntyai import ImageCaptioner

cap = ImageCaptioner()
cap.caption("photo.jpg")
cap.caption("https://example.com/image.jpg")
cap.visualize(["img1.jpg", "img2.jpg", "img3.jpg"])

Ethical Evaluation

from bforbuntyai import EthicalEvaluator

ev = EthicalEvaluator()
ev.evaluate("Some text to check for toxicity and bias")
ev.evaluate_batch(["text one", "text two", "text three"])

Supported Datasets

Dataset	Class	Used by
MNIST	dataset.MNIST()	AutoEncoder, GAN, DCGAN, cGAN, VAE
Fashion-MNIST	dataset.FashionMNIST()	GAN, DCGAN, VAE
CIFAR-10	dataset.CIFAR10()	AutoEncoder
HuggingFace	dataset.HuggingFace("name", split="...")	TextFineTuner
Edges2Shoes	dataset.Edges2Shoes()	Pix2Pix
Custom folder	dataset.Custom("path/", image_size=64)	Any image model

Common API

Every model shares the same interface:

Method	Description
.train(epochs=...)	Train the model, returns self for chaining
.generate(n=...)	Generate samples
.visualize()	Plot results with matplotlib
.save(path)	Save model weights
.load(path)	Load model weights

Cache

Downloaded datasets and models are cached at ~/.bforbuntyai/cache/.