oetu/otis-base-onnx

OTIS base — ONNX

OTIS is an open time series encoder: 7.1 M parameters, matching the accuracy of encoders 54× its size across 162 tasks at 10× less memory, 43× less energy and 37× lower latency. This is the released base checkpoint exported to ONNX so it can run client-side in a browser — the demo page runs it on the visitor's own data without uploading anything.

file	size	contents
otis_encoder.onnx	~22 MB	encoder, float32
otis_decoder.onnx	~5 MB	decoder, float32
otis_encoder_int8.onnx	~6 MB	encoder, dynamic int8
otis_decoder_int8.onnx	~1.4 MB	decoder, dynamic int8
otis_meta.json	—	patch size, domain offsets, variate slots
pos_embed_y.bin	15 KB	learned variate embeddings, raw float32 (39 x 96)
pos_embed_x.bin	17 KB	fixed sincos position table, raw float32 (1 x 43 x 96)
decoder_pos_embed_x.bin	14 KB	the decoder's, raw float32 (1 x 43 x 80)
units_x128_<V>x<T>[_int8].onnx	3-12 MB	UniTS backbone per demo shape
units_meta.json	—	UniTS graph list, token size, prompt count
moment_large_int8.onnx	330 MB	MOMENT-large, dynamic V and T, feat + pred + emb
moment_fc_<dataset>_int8.onnx	3-6 MB	MOMENT's linear-probing forecasting head per benchmark, applied to emb
moment_meta.json	—	MOMENT graph entry
moment_released_tokens.bin, moment_<sine|chirp|square>_tokens.bin	37 KB each	MOMENT's mask token + reconstruction head as graph inputs: the released pair, and the pair fitted on sines, chirps or squares from 20 to 120 Hz for the Impute row
chronos2_fc336_96_int8.onnx	131 MB	Chronos-2, 336 -> 96 median forecast, dynamic V
chronos2_embed_int8.onnx	128 MB	Chronos-2 encoder states for the Explore map
zero_shot_meta.json	—	the zero-shot graph(s) hosted here: split, sizes, int8 error
forecast/<dataset>[_revin]/	~7 MB each	OTIS fine-tuned per forecasting benchmark: int8 graphs, meta, tables
sine_embed.bin, <chirp|square>_embed.bin	384 B each	one OTIS variate embedding per Impute source, fitted on clean sines, chirps or squares from 20 to 120 Hz

Signatures

otis_encoder(x, attn_mask, pos_embed_y, pos_embed_x, ids_keep)                  -> latent
otis_decoder(latent, attn_mask, pos_embed_y, decoder_pos_embed_x, ids_restore)  -> pred

pos_embed_y is (B, V, T', 96) vectors, not integer ids: the nn.Embedding lookup is lifted out of the graph so the caller supplies the variate embeddings. A domain seen in pre-training uses the learned rows from pos_embed_y.bin (indexed arange(V) + 1 + domain_offset); an unseen domain gets randomly initialised ones, trunc_normal(std=0.02), as the paper specifies and as models_otis.initialize_weights does it.

pos_embed_x is (1, 1+T', 96) and is lifted out for the same reason. It is a fixed sincos table, requires_grad=False, so baking the checkpoint's 42 positions in would cap a run at 1008 time steps — which the codebase does not do: util/pos_embed.interpolate_pos_embed_x resamples it (linear, align_corners=False) for anything longer, and slices it for anything shorter. Pass the table you want; pos_embed_x.bin is the checkpoint's.

Masking is deliberately outside the graph: the reference random_masking draws from an RNG and selects with a boolean mask, so its output shape is data dependent and cannot be expressed in ONNX. The caller picks the visible patches and passes them as ids_keep, which also lets a mask-ratio slider re-run without re-exporting. Passing ids_keep = arange(N) with attn_mask = 1 reproduces the all-patches-visible path exactly.

Licenses

The OTIS graphs are released under this repository's license (CC BY-NC 4.0). The comparison models keep their own: UniTS (MIT), MOMENT-large (MIT) and Chronos-2 (Apache-2.0) are converted to ONNX without any change to the weights, and are here only so the demo page can run them in the browser; use them under their own terms. TimesFM-3's license (TimesFM Non-Commercial License v1.0) does not allow the weights or derivatives to be distributed, so its graph is not here: the page runs its rows only from a local copy of the graph.

Note on attention

These weights were trained with an inner nn.MultiheadAttention, i.e. with a second projection between qkv and proj. util/transformer.py keeps that layout and folds the two projections into one at inference, which is exact and is why the exported graph is smaller than the checkpoint's nominal parameter count.