kc-maps / data /splits /README.md
majhas's picture
Checkpoints, reference MD, training splits and generated trajectories
e5b8053 verified
|
Raw History Blame Contribute Delete
2.18 kB

Training splits

These files rebuild <SYSTEM>/train.npz and val.npz from the source datasets.

System Source Pool Train / val / test
AL2 Köhler et al. (2021), AImplicitUnconstrained.tgz, traj0.h5 coordinates 1,000,000 frames in file order 800k / 100k / 100k
AL4 Tan et al. (2025), Ace-AAA-Nme_300K train ++ val ++ test, 700,000 frames 560k / 70k / 70k
AL6 Tan et al. (2025), AAAAAA_310K train ++ val ++ test, 300,000 frames 240k / 30k / 30k

Download the sources (AL2: extract traj0.h5 from the archive above; AL4, AL6: the Sequential Boltzmann Generators data of Tan et al., Hugging Face dataset, under Ace-AAA-Nme_300K/ and AAAAAA_310K/), then, with KCMAPS_DATA the parent data directory, which contains splits/:

hf download transferable-samplers/sequential-boltzmann-generators-data --repo-type dataset --include "Ace-AAA-Nme_300K/*" --local-dir sbg
python -m kcmaps.training.data AL2 AImplicitUnconstrained/traj0.h5
python -m kcmaps.training.data AL4 Ace-AAA-Nme_300K/Ace-AAA-Nme_300K_{train,val,test}.npy
python -m kcmaps.training.data AL6 AAAAAA_310K/AAAAAA_310K_{train,val,test}.npy

Each command checks the sha256 of the split indices and of the Tan et al. files against splits.json, builds the pool in the order above, takes the train and val slices of pool[<SYSTEM>_split_indices.npy] and writes $KCMAPS_DATA/<SYSTEM>/train.npz and val.npz (positions, (frames, atoms, 3) float32 nm). --splits and --out override the two directories.

The indices are

import jax, jax.numpy as jnp, numpy as np
jax.config.update("jax_threefry_partitionable", True)  # JAX 0.9.2
idx = np.asarray(jax.random.permutation(jax.random.PRNGKey(42), jnp.arange(N)))
assert np.array_equal(idx, np.load("AL2_split_indices.npy"))

splits.json lists per system the source files, checksums and counts; the slices follow the counts in train, val, test order.