APINAJA/inverse-fx-order-params-logmel-transformer

Inverse FX Chain Predictor — presence + ORDER + PARAMETERS (three-head Log-Mel Transformer)

From dry + wet audio only, predicts which of EQ / Compressor / Reverb were applied, in which audible order, and with which settings (12 parameters). Fine-tuned from the two-head model APINAJA/inverse-fx-order-logmel-transformer on APINAJA/inverse-fx-phase1-effect-order (revision a5d0eeff7f8d9062b30f1971a2bed342e0de2521), same song-level train / validation / test splits; test used once.

Heads

• presence: P(EQ), P(Compressor), P(Reverb) — BCE on every pair
• order: P(EQ before Compressor), P(Reverb before Compressor) — BCE only where both effects are present
• parameters (new): 12 sigmoid outputs = the dataset's parameters_normalized values (0–1). L1 loss only on the parameters of effects that are really present, averaged per effect. Reverb width is not predicted.

Decoding: most probable of the 12 audibly distinct chains (joint probability), then the parameters of the effects in that chain.

#	chain	presence [EQ,Comp,Rev]	EQ before Comp	Rev before Comp
0	EQ	[1, 0, 0]	—	—
1	Compressor	[0, 1, 0]	—	—
2	Reverb	[0, 0, 1]	—	—
3	EQ + Reverb (order not audible)	[1, 0, 1]	—	—
4	EQ → Compressor	[1, 1, 0]	1	—
5	Compressor → EQ	[1, 1, 0]	0	—
6	Compressor → Reverb	[0, 1, 1]	—	0
7	Reverb → Compressor	[0, 1, 1]	—	1
8	Compressor → {EQ, Reverb}	[1, 1, 1]	0	0
9	EQ → Compressor → Reverb	[1, 1, 1]	1	0
10	Reverb → Compressor → EQ	[1, 1, 1]	0	1
11	{EQ, Reverb} → Compressor	[1, 1, 1]	1	1

#	effect	parameter	unit	scale	range (0 → 1)
0	EQ	frequency	Hz	log	100 – 8000
1	EQ	gain	dB	lin	-12 – 12
2	EQ	q	—	log	0.5 – 4
3	Compressor	threshold	dB	lin	-40 – -15
4	Compressor	ratio	:1	log	2 – 10
5	Compressor	attack	ms	log	1 – 50
6	Compressor	release	ms	log	50 – 500
7	Compressor	makeup_gain	dB	lin	0 – 6
8	Reverb	room_size	—	lin	0.3 – 0.9
9	Reverb	damping	—	lin	0.2 – 0.8
10	Reverb	wet_mix	—	lin	0.15 – 0.5
11	Reverb	dry_mix	—	lin	0.6 – 0.9

lin: value = low + x·(high − low); log: value = low·(high/low)^x.

Fine-tuning

Backbone + presence/order heads copied from the two-head model (incl. log-mel normalisation stats); the parameter head is new. Backbone and old heads are frozen for 1 epoch(s), then trained with LR ×0.3 (old heads ×0.3); parameter head LR 0.0003. Loss = BCE(presence) + 1.0 × BCE(order)

• 1.0 × l1(parameters). Best checkpoint by validation combined_score (combined score = (12-chain accuracy + parameter skill) / 2, parameter skill = 1 − MAE / MAE of the mean-value baseline).

Architecture

dry/wet [B, 441000] → Log-Mel [B, 128, T] each → fusion concat_diff (dry ‖ wet ‖ wet−dry) → [B, T, 384] → Linear → 256 → [CLS] + sinusoidal positions → TransformerEncoder (4 layers, 4 heads, FF 512) → [CLS] → presence [B,3] + order [B,2] + parameters [B,12]. 2,309,137 trainable parameters.

Results — test (best-validation checkpoint, epoch 21)

metric	model	trivial baseline
12-chain accuracy	0.5394	0.1333 (majority class)
order accuracy (both questions)	0.6629	0.5 (coin flip)
presence macro F1	0.9323	0.8462 (always "present")
parameter MAE (0-1, macro over 12)	0.2133	0.2541 (predict training mean)
parameter skill	0.160	0
parameters within tolerance	0.412	—
combined score	0.3499	—

Parameter errors are measured for effects that are really present (given the true chain).

parameter	n	MAE (0-1)	mean-baseline MAE	typical error	baseline error	within tolerance	baseline	R²
EQ frequency	726	0.1508	0.2397	0.95 oct	1.52 oct	38.7%	16.1%	0.442
EQ gain	726	0.1771	0.3149	4.25 dB	7.56 dB	24.1%	0.0%	0.495
EQ Q	726	0.1954	0.2484	×1.50	×1.68	59.8%	38.3%	0.233
Comp threshold	726	0.1398	0.2316	3.50 dB	5.79 dB	51.2%	26.9%	0.572
Comp ratio	726	0.2205	0.2477	×1.43	×1.49	60.5%	51.0%	0.149
Comp attack	726	0.2266	0.2420	×2.43	×2.58	44.5%	38.7%	0.050
Comp release	726	0.2571	0.2601	×1.81	×1.82	35.1%	31.8%	-0.051
Comp makeup	726	0.2305	0.2528	1.38 dB	1.52 dB	37.6%	34.3%	0.113
Rev room size	726	0.2245	0.2391	0.13	0.14	39.1%	35.0%	0.048
Rev damping	726	0.2666	0.2601	0.16	0.16	29.9%	28.7%	-0.101
Rev wet mix	726	0.2237	0.2516	0.08	0.09	36.9%	30.9%	0.122
Rev dry mix	726	0.2474	0.2609	0.07	0.08	37.3%	32.8%	0.015

effect	precision	recall	F1
EQ	0.9378	0.8306	0.8809
Compressor	0.9986	0.9683	0.9832
Reverb	0.9592	0.9077	0.9328

Validation at the best epoch: 12-chain 0.5700, parameter MAE 0.2211, combined score 0.3593.

Training history

epoch	train loss	val loss	val 12-chain acc	val param MAE	val param skill	val combined	best
001	0.7200	1.3152	0.5043	0.2461	0.053	0.2785	yes
002	0.7182	1.3278	0.4889	0.2385	0.082	0.2853	yes
003	0.7351	1.3206	0.4831	0.2382	0.083	0.2830
004	0.7187	1.2864	0.5208	0.2323	0.106	0.3132	yes
005	0.7063	1.2386	0.5237	0.2282	0.121	0.3226	yes
006	0.6844	1.3273	0.5082	0.2285	0.120	0.3143
007	0.6686	1.3015	0.5237	0.2289	0.119	0.3213
008	0.6594	1.3409	0.5159	0.2264	0.128	0.3221
009	0.6431	1.3225	0.5256	0.2248	0.135	0.3301	yes
010	0.6320	1.2759	0.5295	0.2257	0.131	0.3302	yes
011	0.6190	1.3604	0.5304	0.2230	0.142	0.3360	yes
012	0.6148	1.3678	0.5169	0.2251	0.133	0.3251
013	0.5974	1.3943	0.5188	0.2244	0.136	0.3275
014	0.5937	1.3454	0.5391	0.2226	0.143	0.3411	yes
015	0.5864	1.3856	0.5382	0.2225	0.143	0.3407
016	0.5769	1.3767	0.5498	0.2211	0.149	0.3492	yes
017	0.5665	1.3991	0.5469	0.2207	0.150	0.3486
018	0.5652	1.4136	0.5633	0.2216	0.147	0.3550	yes
019	0.5547	1.4574	0.5527	0.2223	0.144	0.3485
020	0.5523	1.4080	0.5536	0.2209	0.149	0.3516
021	0.5465	1.3828	0.5700	0.2211	0.149	0.3593	yes
022	0.5458	1.4251	0.5536	0.2210	0.149	0.3513
023	0.5443	1.4168	0.5556	0.2210	0.149	0.3524
024	0.5398	1.4237	0.5643	0.2211	0.149	0.3565
025	0.5494	1.4256	0.5565	0.2211	0.149	0.3527

Files

final/ (model.pt, config.json, label_map.json, test_metrics.json), best/, checkpoint_epoch_XXX/ (full resumable state), results/ (history, plots, test predictions incl. parameters), runs/ (TensorBoard), fxp3/ (model code), train.py.