Skip to content

Outputs

AdaptivePy writes structured outputs to output_dir for reproducibility and downstream analysis.

Top-level artifacts

File Description
run_config.yaml Exact copy of the run configuration
assignments.npy Per-frame cluster labels (1D integer array)
cluster_model.pkl Serialized clustering model (joblib)
metadata.csv Cluster populations (cluster_id, population)
logs.txt Full run log
combined_metadata.csv Seeds from all policies (multi-policy runs only)

Clustering artifacts (assignments.npy, cluster_model.pkl, top-level metadata.csv) are omitted when all configured policies are frame-level, such as maxent_vampnet or ts_dar alone.

Per-policy outputs

Each policy gets a subdirectory named after the policy:

results/least_counts/
├── seeds.csv
├── metadata.csv
└── pdbs/          # optional, when trajectories are provided
    └── seed_0_traj0_frame42.pdb

seeds.csv

Column Description
seed_id Sequential ID within the policy
policy Policy name
traj_id Source trajectory index
frame_id Frame index within the trajectory
cluster_id Cluster the seed was drawn from (blank for frame-level policies)
global_index Row index in the concatenated feature matrix

Example:

seed_id,policy,traj_id,frame_id,cluster_id,global_index
0,least_counts,0,42,3,42
1,least_counts,1,15,1,65

metadata.csv

Cluster population statistics (same format as the top-level file):

cluster_id,population
0,123
1,45
2,67

FAST sidecar files

When using the fast policy, fast/scores.csv is written with columns: cluster_id, directed_score, exploration_score, reward, population.

MA-REAP sidecar files

When using the ma_reap policy, additional CSV files are written under ma_reap/:

File Description
scores.csv Per-candidate aggregate_score, population, and score_{agent} columns
agent_weights.csv Learned CV weights: agent, feature_index, weight
stakes.csv Agent stakes per candidate: cluster_id, stake_{agent} columns
executors.csv Seed executor assignment: seed_id, cluster_id, executor_agent

Example ma_reap/executors.csv:

seed_id,cluster_id,executor_agent
0,3,agent_0
1,7,agent_1

MA-REAP requires assigning every feature trajectory to an agent via policy_params.ma_reap.agents. See Features and Policies.

MaxEnt VAMPNet sidecar files

When using the maxent_vampnet policy, maxent_vampnet/scores.csv is written with per-frame entropy and softmax probabilities:

Column Description
global_index Row index in the concatenated feature matrix
traj_id Source trajectory index
frame_id Frame index within the trajectory
entropy Shannon entropy of softmax state probabilities
selected Whether the frame was chosen as a seed
prob_0 … prob_K Softmax metastable-state probabilities

Frame-level MaxEnt runs do not write per-policy metadata.csv.

TS-DAR sidecar files

When using the ts_dar policy, ts_dar/scores.csv is written with per-frame OOD scores, assigned states, embeddings, and softmax probabilities:

Column Description
global_index Row index in the concatenated feature matrix
traj_id Source trajectory index
frame_id Frame index within the trajectory
ood_score TS-DAR out-of-distribution score
state Softmax state assignment
selected Whether the frame was chosen as a seed
emb_0 … emb_D Hyperspherical embedding coordinates
prob_0 … prob_K Softmax metastable-state probabilities

Frame-level TS-DAR runs do not write per-policy metadata.csv.

Metapolicy outputs

When metapolicy.enabled: true, AdaptivePy writes the final ensemble seed set under metapolicy/:

File Description
seeds.csv Final ensemble seeds; the policy column uses metapolicy.name
metadata.csv Cluster populations for the clustered ensemble candidates
votes.csv Cluster-level audit table with selection status, vote count, ensemble score, per-policy ranks, and per-policy scores

Metapolicy output is cluster-level. If MaxEnt VAMPNet or TS-DAR is included, its frame entropy or OOD score is aggregated to clusters before voting or allocation.

PDB export

When trajectories_dir, topology, and write_pdbs: true are set, selected seed frames are extracted with mdtraj and saved as PDB files under pdbs/.

Filename pattern:

seed_{seed_id}_traj{traj_id}_frame{frame_id}.pdb

Reproducibility

Every run preserves:

  • Configuration snapshot
  • Random seed
  • Cluster model and assignments
  • Full metadata and logs

Re-run with the same config and seed to reproduce identical cluster assignments (for deterministic clustering methods).

See also