CoolFace
Datasetpublic

chansigit/transmap-layouts

TransMap gene layouts Canonical gene-to-pixel layouts used by the TransMap single-species and multi-species experiments. Single-species layouts The repository contains 65 canonical HDF5 layouts for five datasets: ds_lung ds_breast ds_cite_seq_pbmc ds_gut_fetal ds_lung_fetal_donor Each dataset has a 48 x 48 layout over 2,000 HVGs for four layout families: one TorchGW layout: layout_torchgw_snn_k7_48x48.h5 one IGTD layout: layout_igtd_linear_abs_seed1_48x48.h5… See the full description on the dataset page: https://huggingface.co/datasets/chansigit/transmap-layouts.

sourceHugging Faceupdated 2mo agoView on Hugging Face
0likes26downloads
Dataset Card

TransMap gene layouts

Canonical gene-to-pixel layouts used by the TransMap single-species and multi-species experiments.

Single-species layouts

The repository contains 65 canonical HDF5 layouts for five datasets:

  • ds_lung
  • ds_breast
  • ds_cite_seq_pbmc
  • ds_gut_fetal
  • ds_lung_fetal_donor

Each dataset has a 48 x 48 layout over 2,000 HVGs for four layout families:

  • one TorchGW layout: layout_torchgw_snn_k7_48x48.h5
  • one IGTD layout: layout_igtd_linear_abs_seed1_48x48.h5
  • one DeepInsight layout: layout_deepinsight_tsne_lsa_seed0_48x48.h5
  • ten random layouts: layout_random_seed{0..9}_48x48.h5

The files preserve the training data directory convention: ds_<dataset>/gene_layout/<layout>.h5.

Multi-species layout

The historical five-species layout used by the genofoundation multi-species training runs is stored under:

multi_species/string_5species/headline_tmax8_seed42/

It contains one HDF5 file for each species:

  • homo_sapiens
  • mus_musculus
  • danio_rerio
  • drosophila_melanogaster
  • caenorhabditis_elegans

The HDF5 metadata records a 144 x 144 grid, T_max=8, seed 42, hard anchoring, and the D4 canonical transform flip_v. This is a byte-for-byte archive of the historical training artifact; it was not rebuilt during the TransMap migration. snapshot.json and SHA256SUMS record provenance and integrity metadata.

This artifact differs from the currently migrated build recipe, which requests T_max=5. Keep the snapshot and regenerated layouts under distinct names.

Download

Download the full dataset with huggingface_hub:

python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="chansigit/transmap-layouts",
    repo_type="dataset",
    local_dir="transmap-layouts",
)

Or select only the historical multi-species snapshot:

python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="chansigit/transmap-layouts",
    repo_type="dataset",
    allow_patterns="multi_species/string_5species/headline_tmax8_seed42/*",
    local_dir="transmap-layouts",
)

Source code

Build and migration workflows live in `chansigit/transmap-layout-builds`.