Skip to content

Model reference

Generated from inmotion.models.registry. inmotion model verify proves these are loadable; inmotion model list prints the same table.

features is what the model consumes at its input: raw is the 10 RSSI values, rich is the 18-channel expansion (raw, first and second differences, window deviation, spectral and shape summaries).

params is the exact module parameter count; paper is the figure printed in the paper’s table. Where the two disagree the stored checkpoint is not the model that table describes, and the gap is recorded in PROVENANCE_GAPS.

Canonical model count: 18

name factory features params paper (M) checkpoint
cf_jepa cf_jepa rich 1,232,261 1.232 20-aug/models/exotic/normal-new-ds/full-after-hpo/cf_jepa_ft_seed42.pt
cf_jepa_base cf_jepa rich 2,153,349 - 20-aug/models/exotic/normal-new-ds/cf_jepa_ft_seed42.pt
deep_stack deep_stack raw 43,575,549 43.576 backup/DeepStackEnsemble_seed42.pt
lejepa lejepa raw 565,765 0.566 20-aug/models/exotic/normal-new-ds/full-after-hpo/lejepa_ft_seed42.pt
lejepa_base lejepa rich 2,215,813 - 20-aug/models/exotic/normal-new-ds/lejepa_ft_seed42.pt
mamba3_cnn mamba3_cnn raw 945,028 - backup/mamba3_cnn_seed42.pt
mamba3_multiview mamba3_multiview raw 1,193,349 - backup/mamba3_multiview_seed42.pt
mamba3_tcn mamba3_tcn raw 695,812 - backup/mamba3_tcn_seed42.pt
mamba3_transformer mamba3_transformer raw 561,796 - backup/mamba3_transformer_seed42.pt
sigreg sigreg raw 696,900 0.697 20-aug/models/exotic/dataset-csv/sigreg_seed42.pt
sigreg_after_hpo sigreg raw 317,460 - 20-aug/models/exotic/normal-new-ds/full-after-hpo/sigreg_seed42.pt
sigreg_base sigreg raw 474,564 - 20-aug/models/exotic/normal-new-ds/sigreg_seed42.pt
sigreg_icaisf sigreg raw 696,900 0.697 backup/sigreg_seed42.pt
sigreg_seed3 sigreg raw 696,900 0.697 backup/sigreg_seed3.pt
sigreg_seed5 sigreg raw 696,900 0.697 backup/sigreg_seed5.pt
t_jepa t_jepa rich 633,477 0.633 20-aug/models/exotic/normal-new-ds/full-after-hpo/t_jepa_ft_seed42.pt
t_jepa_base t_jepa rich 2,347,653 - 20-aug/models/exotic/normal-new-ds/t_jepa_ft_seed42.pt
ts_jepa ts_jepa rich 2,153,349 2.153 20-aug/models/exotic/normal-new-ds/ts_jepa_ft_seed42.pt
  • factory: cf_jepa
  • checkpoint: 20-aug/models/exotic/normal-new-ds/full-after-hpo/cf_jepa_ft_seed42.pt
  • state-dict prefix: -
  • architecture: (factory defaults)
  • note: after-HPO frontier point
  • factory: cf_jepa
  • checkpoint: 20-aug/models/exotic/normal-new-ds/cf_jepa_ft_seed42.pt
  • state-dict prefix: -
  • architecture: dim_feedforward=512, num_layers=4
  • note: pre-HPO base run (2.1)
  • factory: deep_stack
  • checkpoint: backup/DeepStackEnsemble_seed42.pt
  • state-dict prefix: -
  • architecture: (factory defaults)
  • note: 10-base deep stack; the published table cites this under the stale path models/dl/augmented/DeepStackEnsemble_seed42.pt
  • provenance gap: The published table points at models/dl/augmented/DeepStackEnsemble_seed42.pt, which does not exist. The equivalent artifact used here is backup/DeepStackEnsemble_seed42.pt.
  • factory: lejepa
  • checkpoint: 20-aug/models/exotic/normal-new-ds/full-after-hpo/lejepa_ft_seed42.pt
  • state-dict prefix: -
  • architecture: d_model=128, dim_feedforward=256, nhead=4, num_layers=4, pred_num_layers=3, sigreg_lambda=0.05
  • note: after-HPO Pareto winner
  • factory: lejepa
  • checkpoint: 20-aug/models/exotic/normal-new-ds/lejepa_ft_seed42.pt
  • state-dict prefix: -
  • architecture: d_model=256, dim_feedforward=512, in_channels=18, nhead=8, num_layers=4, pred_num_layers=2, sigreg_lambda=0.1
  • note: pre-HPO base run (2.1)
  • factory: mamba3_cnn
  • checkpoint: backup/mamba3_cnn_seed42.pt
  • state-dict prefix: -
  • architecture: (factory defaults)
  • note: experimental; not in the reported results
  • factory: mamba3_multiview
  • checkpoint: backup/mamba3_multiview_seed42.pt
  • state-dict prefix: -
  • architecture: d_model=256, d_state=16, mimo_rank=4, n_mamba_layers=2
  • note: experimental; not in the reported results
  • factory: mamba3_tcn
  • checkpoint: backup/mamba3_tcn_seed42.pt
  • state-dict prefix: -
  • architecture: d_model=128, d_state=16, mimo_rank=4, n_mamba_layers=2, tcn_channels=128, tcn_depth=4
  • note: experimental; not in the reported results
  • factory: mamba3_transformer
  • checkpoint: backup/mamba3_transformer_seed42.pt
  • state-dict prefix: -
  • architecture: d_model=128, d_state=16, mimo_rank=4, nhead=4, num_blocks=2
  • note: experimental; not in the reported results
  • factory: sigreg
  • checkpoint: 20-aug/models/exotic/dataset-csv/sigreg_seed42.pt
  • state-dict prefix: m.
  • architecture: num_filters=192
  • note: retrained on dataset.csv: 192 filters / 3 blocks / latent 128, seed 42, lr 1e-3, batch 128, sigreg_lambda 0.01, 29 of 150 epochs, val MCC 0.6181, test MCC 0.6333; reproduce with inmotion train exotic --model sigreg --data data/processed/dataset.csv --seed 42, whose defaults ARE this architecture (–num-filters 192, –num-blocks 3, –latent-dim 128); full config in 20-aug/models/exotic/dataset-csv/sigreg_seed42.json
  • factory: sigreg
  • checkpoint: 20-aug/models/exotic/normal-new-ds/full-after-hpo/sigreg_seed42.pt
  • state-dict prefix: m.
  • architecture: num_filters=128
  • note: 128-filter retrain at the HPO winner config (2.2)
  • factory: sigreg
  • checkpoint: 20-aug/models/exotic/normal-new-ds/sigreg_seed42.pt
  • state-dict prefix: m.
  • architecture: num_blocks=2, num_filters=192
  • note: pre-HPO base run (2.1), 2 residual blocks
  • factory: sigreg
  • checkpoint: backup/sigreg_seed42.pt
  • state-dict prefix: m.
  • architecture: num_filters=192
  • note: superseded by sigreg for dataset.csv work; this is the artifact the published table cites (192-filter encoder, icaisf split)
  • factory: sigreg
  • checkpoint: backup/sigreg_seed3.pt
  • state-dict prefix: m.
  • architecture: num_filters=192
  • note: multi-seed diversity member
  • factory: sigreg
  • checkpoint: backup/sigreg_seed5.pt
  • state-dict prefix: m.
  • architecture: num_filters=192
  • note: multi-seed diversity member
  • factory: t_jepa
  • checkpoint: 20-aug/models/exotic/normal-new-ds/full-after-hpo/t_jepa_ft_seed42.pt
  • state-dict prefix: -
  • architecture: d_model=128, dim_feedforward=256, nhead=4, num_layers=4, pred_dim=64, pred_num_layers=3
  • note: after-HPO Pareto winner
  • factory: t_jepa
  • checkpoint: 20-aug/models/exotic/normal-new-ds/t_jepa_ft_seed42.pt
  • state-dict prefix: -
  • architecture: d_model=256, dim_feedforward=512, nhead=8, num_layers=4, pred_dim=128, pred_num_layers=2
  • note: pre-HPO base run (2.1); nhead/pred_dim do not affect any saved tensor shape, so they are not recoverable from the checkpoint
  • factory: ts_jepa
  • checkpoint: 20-aug/models/exotic/normal-new-ds/ts_jepa_ft_seed42.pt
  • state-dict prefix: -
  • architecture: (factory defaults)
  • note: base run (18ch, embed 256)

Recorded rather than resolved. tests/test_checkpoint_contract.py fails if one of these is closed by editing the paper figure or the checkpoint count without restoring the matching artifact.

deep_stack

The published table points at models/dl/augmented/DeepStackEnsemble_seed42.pt, which does not exist. The equivalent artifact used here is backup/DeepStackEnsemble_seed42.pt.