3. Skip to content

3. Optional extras and platform support

Use this guide to choose the smallest dependency set that matches your workflow. Extras in ModSSC are grouped by provider family, preprocess backends, training backends, and meta bundles.

3.1 Quick recommendations

Goal Install
Core package API and CLI only python -m pip install modssc
Repository workflows with benchmarks, examples, and notebooks python -m pip install -e "."
Most dataset providers in one shot python -m pip install "modssc[datasets]"
Most preprocessing backends in one shot python -m pip install "modssc[preprocess]"
Full local feature set python -m pip install "modssc[full]"
Docs authoring python -m pip install "modssc[docs]"
Development and tests python -m pip install "modssc[dev]"

3.2 Provider and dataset extras

Extra Unlocks Typical use
openml OpenML-backed tabular datasets iris, adult, breast_cancer
hf Hugging Face datasets provider ag_news, imdb, amazon_polarity
text Alias used for text dataset installs Text datasets backed by datasets
tfds TensorFlow Datasets provider TFDS-backed datasets
vision torchvision-backed datasets mnist, cifar10, svhn, stl10
audio torchaudio-backed datasets speechcommands, yesno
graph PyG-backed graph datasets and graph-oriented tooling cora, citeseer, pubmed
datasets / data Meta bundle for common dataset providers Broad data-loading setup

3.3 Method and model extras

Extra Unlocks Typical use
inductive-torch Torch-based inductive methods deep inductive SSL workflows
inductive-tf TensorFlow-based inductive methods TF inductive experiments
supervised-torch Torch supervised baselines MLP/CNN/LSTM-style baselines
supervised-torch-geometric torch-geometric supervised baselines GraphSAGE inductive baseline
transductive-torch Torch-based transductive methods GCN, GAT, APPNP and similar
transductive-pyg PyG-backed transductive support graph-native torch workflows
transductive-advanced Advanced transductive torch/PyG stack larger graph method surface
graph-faiss FAISS graph construction backend fast approximate kNN on supported platforms

3.4 Preprocess extras

Extra Unlocks Typical use
preprocess-sklearn scikit-learn preprocess steps imputation, scaling, PCA
preprocess-text sentence-transformers and transformers text embeddings
preprocess-vision torchvision, Pillow, OpenCLIP image transforms and pretrained encoders
preprocess-audio torchaudio preprocess stack waveform/audio features
preprocess-graph scipy-based graph preprocess steps sparse graph features
preprocess Meta bundle for preprocess extras mixed preprocess workflows

3.5 Platform notes

  • The benchmark runner and authored benchmark configs live in bench/ and are repository assets, not PyPI package assets.
  • graph-faiss is guarded by a platform marker and does not install on Darwin in pyproject.toml.
  • full is the easiest way to get a rich local environment, but it is also the heaviest dependency profile.
  • If you only need one workflow, prefer the narrow extra over full so environments stay smaller and easier to maintain.

3.6 Practical install recipes

CPU-friendly local benchmark setup:

python -m pip install -e "."
python -m pip install "modssc[datasets,preprocess,inductive-torch,transductive-torch]"

Text-oriented inductive workflow:

python -m pip install "modssc[hf,preprocess-text,inductive-torch]"

Graph-oriented transductive workflow:

python -m pip install "modssc[graph,transductive-torch]"
Sources
  1. pyproject.toml
  2. src/modssc/data_loader/providers/
  3. src/modssc/preprocess/steps/
  4. src/modssc/supervised/registry_data.py
  5. bench/README.md