amsks/psmflows-preimages

PSMFlows noise preimages

Precomputed latents for the PSMFlows pipeline (code). A behaviour-cloned conditional flow G(s, u) maps Gaussian noise to dataset actions; these files store, for every transition of an OGBench dataset, the noise u that decodes to the recorded action. Computing them takes 4-19 h on a GPU per environment; downloading them takes minutes.

Contents

preimages/<env>.npz              preimage-augmented dataset (training input)
preimages/<env>.npz.meta.json    provenance: env, Stage-A checkpoint, inversion config
flow/<env>/params_500000.pkl     the frozen behaviour flow the latents invert
flow/<env>/flags.json            the behaviour-flow training config

Each .npz holds the OGBench transitions (observations, actions, rewards, terminals, masks, next_observations) plus:

key	shape	what it is
noise_preimage_point	(N, d_a)	backward-ODE point preimage u*, G(s, u*) ~ a
noise_preimage_mean	(N, K, d_a)	Gaussian mixture posterior over u, means
noise_preimage_cov	(N, K, d_a, d_a)	mixture covariances
noise_preimage_weights	(N, K)	mixture weights
preimage_ess	(N,)	effective sample size of the final EM iterate (of 200)
preimage_roundtrip	(N,)	|G(s, u*) - a|, the inversion residual
preimage_valid	(N,)	1.0 if every preimage product for the row is finite

Quality

env	rows	d_a	mean |u*|^2 (expected)	mean ESS	frac ESS>20	median round-trip	invalid rows
pointmaze-medium-navigate	1,000,000	2	1.64 (2)	94.6	0.90	5.5e-05	0
cube-single-play	1,000,000	5	5.59 (5)	81.3	0.84	1.2e-04	13
antmaze-medium-navigate	1,000,000	8	8.49 (8)	7.6	0.06	2.3e-04	881

The point preimages are healthy in every environment: \|u*\|^2 sits near its expected value d_a under the standard normal prior, and the round-trip residual is ~1e-4.

The antmaze mixture posterior is not usable. Its mean ESS is far below the rest and only a few percent of rows clear ESS > 20; alpha=20 was tuned at d_a=5 and does not transfer to d_a=8. Use agent.use_point_preimage=true there (which is what our own antmaze runs use), or re-run the inversion with a re-tuned alpha.

Inversion settings

{
  "pointmaze-medium-navigate": {
    "num_clusters": 1,
    "alpha": 20.0,
    "num_samples": 200,
    "n_steps": 10,
    "n_initial_steps": 100,
    "batch_size": 256,
    "seed": 0,
    "allow_untrained": false
  },
  "cube-single-play": {
    "num_clusters": 1,
    "alpha": 20.0,
    "num_samples": 200,
    "n_steps": 10,
    "n_initial_steps": 100,
    "batch_size": 256,
    "seed": 0,
    "allow_untrained": false
  },
  "antmaze-medium-navigate": {
    "num_clusters": 1,
    "alpha": 20.0,
    "num_samples": 200,
    "n_steps": 10,
    "n_initial_steps": 100,
    "batch_size": 256,
    "seed": 0,
    "allow_untrained": false
  }
}

n_initial_steps must equal the flow's flow_steps and both must be >= 100: the implicit Euler inverse diverges at the training default of 10.

Use

from huggingface_hub import hf_hub_download

npz = hf_hub_download('amsks/psmflows-preimages', 'preimages/cube-single-play.npz', repo_type='dataset')

Then train against the matching flow checkpoint; see docs/PREIMAGES.md.

Provenance

The transition arrays are copied from OGBench datasets and are redistributed here only so that a .npz is a drop-in training input; OGBench is the original source. The preimage arrays and flow checkpoints are ours.