Encoder cards

An encoder card is a YAML mapping with one key, the encoder id. Cards of the released encoders are in bench/registry/encoders.yaml; cards registered with imu-bench eval --encoder <file> go to bench/registry/encoders.d/<id>.yaml.

Fields

Field	Required	Meaning
name	yes	display name, with the reference
params_m	yes	parameter count in millions
pretrain_data	yes	the pretraining corpus, dataset by dataset with the split used (audited against the evaluation sets)
normalize_method	yes	zscore (per channel and window), physical (m/s² and rad/s as stored) or raw; applied by the harness before the encoder, identically in every cell
channels	yes	6 (accelerometer + gyroscope) or 3 (accelerometer only)
track	yes	open or controlled
external_encoder	plugin / adapters	plugin for the plugin interface; harnet, normwear, inertia1 … for the built-in adapters
entry	plugin	package.module:build; build(weights=<path or None>, **kwargs) returns an nn.Module with an integer embed_dim
weights	plugin	local path, hf://<org>/<repo>/<file> (downloaded with huggingface_hub), or builtin
kwargs	plugin	keyword arguments passed to build
model_hz	plugin	sampling rate the model expects; the harness resamples from 50 Hz
acc_units	plugin	m/s2 (default) or g
weights_url	open track	where the weights are published
run, ckpt	built-in	run directory under runs/ and checkpoint name (best, latest, …) for encoders trained with this repository's trainer
alt_attention, shared_kernel, use_cls	built-in	architecture flags of the shared ViT1D
preprocess_from	built-in	pretraining config whose preprocessing the cells inherit
group, order, short, hide_in_summary	optional	display metadata for the tables

Plugin contract

def build(weights=None, **kwargs) -> torch.nn.Module   # attribute embed_dim: int
module(x)   # x: [B, C, T] float at model_hz; acc x/y/z then gyro x/y/z, wrist FLU frame
            # -> {"tokens": [B, N, D], "pooled": [B, D]}, or only [B, N, D], or only [B, D]

tokens are read by the pose and odometry decoders on the harness grid of T // 16 patches (0.32 s at 50 Hz); pooled is read by the classification and retrieval heads. A model that returns only a pooled vector is repeated over the token grid; a model that returns only tokens is mean-pooled. See examples/plugin/inrtlwrist_plugin.py.