add compression to property files

[sfluegel05]

Property files take up an absurd amount of disk space.

For instance, for a ChEBI50 dataset (v244), the graph_properties\properties folder has nearly 10GB. AtomType_one_hot.pt alone has 6.4GB.

This PR compresses data before saving it to a file (and decompresses when loading). Especially the one-hot encodings don't need to be saved as one-hot tensors. So they are now stored as 1-d and re-converted to one-hot when loading.

File sizes are reduced by a factor of up to 109:

DONE  AtomAromaticity_bool.pt: 106.9 MB -> 59.6 MB (1.8x)
DONE  AtomCharge_one_hot.pt: 743.5 MB -> 59.0 MB (12.6x)
DONE  AtomHybridization_one_hot.pt: 427.6 MB -> 61.3 MB (7.0x)
DONE  AtomNumHs_one_hot.pt: 424.4 MB -> 58.7 MB (7.2x)
DONE  AtomType_one_hot.pt: 6376.0 MB -> 58.5 MB (109.0x)
DONE  BondAromaticity_bool.pt: 109.5 MB -> 59.8 MB (1.8x)
DONE  BondInRing_bool.pt: 107.1 MB -> 58.2 MB (1.8x)
DONE  BondType_one_hot.pt: 331.8 MB -> 58.5 MB (5.7x)
DONE  NumAtomBonds_one_hot.pt: 639.1 MB -> 59.6 MB (10.7x)
DONE  RDKit2DNormalized_asis.pt: 360.7 MB -> 203.5 MB (1.8x)

[Copilot]

Pull request overview

This PR introduces a compression/decompression layer for cached molecular property tensors to substantially reduce on-disk storage, with special handling to avoid storing large dense one-hot matrices.

Changes:

• Add compress() / decompress() hooks to PropertyEncoder, with concrete implementations for one-hot, boolean, and float “as-is” encoders.
• Apply compression when saving processed property caches and decompression immediately after loading them in the ChEBI dataset pipeline.
• Store one-hot encodings as compact class-index vectors on disk and reconstruct one-hot tensors at load time.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
chebai_graph/preprocessing/property_encoder.py	Adds encoder-level compression/decompression APIs and implements compact storage for one-hot/bool/float encodings.
chebai_graph/preprocessing/datasets/chebi.py	Integrates compression into property caching (torch.save) and decompression into property loading (torch.load).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.