OpenVino Training Kit for Scikit-Learn and Pytorch

[leoheim]

Summary

Mentor: '@SHIVAMBASIA'

This PR adds comprehensive scikit-learn and PyTorch model wrappers with OpenVINO optimization for the GSoC 2025 "AI PC Model Training Kit" project. These wrappers enable developers to train, evaluate, and deploy classical ML and deep learning models locally using OpenVINO's optimized backend.

Key features:

• scikit-learn wrappers for classification, regression, clustering, decomposition, and neighbors
• PyTorch wrappers for classification, regression, segmentation, and detection
• Automatic OpenVINO optimization via sklearnex and Intel Extension for PyTorch (IPEX)
• Model persistence: joblib for scikit-learn, checkpointing for PyTorch
• OpenVINO IR export for supported models

Supported models:

• scikit-learn: LogisticRegression, RandomForestClassifier, KNeighborsClassifier, SVC, NuSVC, LinearRegression, Ridge, ElasticNet, Lasso, RandomForestRegressor, KNeighborsRegressor, SVR, NuSVR, KMeans, DBSCAN, PCA, IncrementalPCA, TSNE, NearestNeighbors
• PyTorch: Any nn.Module (e.g., ResNet, UNet) via wrappers for classification, regression, segmentation, detection

---

How to test

scikit-learn example

from ov_training_kit.sklearn import LogisticRegression
from sklearn.datasets import make_classification

X, y = make_classification(n_samples=1000, n_features=20, random_state=42)
model = LogisticRegression()
model.fit(X, y)
model.save_model("test_model.joblib")
model.load_model("test_model.joblib")
model.convert_to_ir(X, model_name="test_model")

PyTorch example

import torch
from torchvision.models import resnet18
from ov_training_kit.pytorch import ClassificationWrapper

model = resnet18(pretrained=True)
model.fc = torch.nn.Linear(model.fc.in_features, 10)
wrapper = ClassificationWrapper(model)

X = torch.randn(32, 3, 224, 224)
y = torch.randint(0, 10, (32,))
loader = torch.utils.data.DataLoader(list(zip(X, y)), batch_size=8)

from torch import nn, optim
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(wrapper.model.parameters())
wrapper.train(loader, criterion, optimizer, num_epochs=1, device="cpu")

wrapper.compile()

def accuracy_metric(preds, targets):
    return (preds.argmax(dim=1) == targets).float().mean().item()
score = wrapper.evaluate(loader, accuracy_metric, device="cpu")
print("Accuracy:", score)

See the README files in src/sklearn and src/pytorch for more details and usage examples.

[SHIVAMBASIA]

@alvoron Can you please review the code and provide your feedback?

[alvoron]

1. Please update https://github.com/openvinotoolkit/openvino_contrib/blob/master/README.md
2. Please add all 3rd party dependencies to https://github.com/openvinotoolkit/openvino_contrib/blob/master/third-party-programs.txt

[leoheim]

@alvoron I've updated the Pr with the changes, thanks!