Inverse FX Chain Predictor — effect presence + ORDER (two-head Log-Mel Transformer)
From dry + wet audio only, predicts which of EQ / Compressor / Reverb were applied and in which audible order.
Trained from scratch on APINAJA/inverse-fx-phase1-effect-order
(revision a5d0eeff7f8d9062b30f1971a2bed342e0de2521), song-level train / validation / test splits; test used once.
Why two heads (and 12, not 15, chains)
EQ (peaking biquad) and Reverb (Freeverb) are linear, so EQ→Reverb and Reverb→EQ give the same audio. Only the Compressor's position is audible. The model therefore predicts
- presence head: P(EQ), P(Compressor), P(Reverb) — BCE on every pair
- order head: P(EQ before Compressor), P(Reverb before Compressor) — BCE only where both effects are present
and decodes the most probable of the 12 audibly distinct chains (joint probability, always a valid chain). The dataset's 15 chains map onto them; a perfect model can reach at most 80 % 15-chain accuracy.
| # | 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 |
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 head [B,3] + order head [B,2].
2,274,693 trainable parameters. Loss = BCE(presence) + 1.0 × masked BCE(order).
Preprocessing: 44100 Hz mono, 10.0 s, n_fft 2048, hop 512,
128 mels 20.0–20000.0 Hz, log-mel standardised (mean -6.1873, std 4.7023).
Results — test (best-validation checkpoint, epoch 27, selected by validation chain12_accuracy)
| metric | model | trivial baseline |
|---|---|---|
| 12-chain accuracy (main) | 0.4717 | 0.1333 (majority class) |
| 15-chain accuracy | 0.3657 | ceiling 0.8000 for a perfect model |
| order accuracy (both questions) | 0.6458 | 0.5 (coin flip) |
| EQ before Compressor | 0.5568 | 0.5 |
| Reverb before Compressor | 0.7348 | 0.5 |
| presence macro F1 | 0.9141 | 0.8462 (always "present") |
| presence exact match | 0.6677 | — |
| effect | precision | recall | F1 |
|---|---|---|---|
| EQ | 0.8714 | 0.8030 | 0.8358 |
| Compressor | 1.0000 | 0.9711 | 0.9853 |
| Reverb | 0.9502 | 0.8939 | 0.9212 |
Order accuracy vs audibility (test)
| audibility | accuracy | n |
|---|---|---|
| < −40 dB (practically identical) | 0.5208 | 48 |
| −40 … −20 dB (subtle) | 0.5607 | 280 |
| > −20 dB (clear) | 0.6868 | 728 |
Validation at the best epoch: 12-chain 0.5043, order 0.6766, presence macro F1 0.9207.
Training history
| epoch | train loss | val loss | val 12-chain acc | val order acc | val presence F1 | best |
|---|---|---|---|---|---|---|
| 001 | 1.2473 | 1.1621 | 0.2290 | 0.5580 | 0.8563 | yes |
| 002 | 1.0799 | 1.0179 | 0.3285 | 0.5534 | 0.9033 | yes |
| 003 | 1.0045 | 0.9887 | 0.3478 | 0.5888 | 0.9124 | yes |
| 004 | 0.9639 | 0.9641 | 0.3787 | 0.6150 | 0.9121 | yes |
| 005 | 0.9362 | 0.9385 | 0.3594 | 0.6024 | 0.9103 | |
| 006 | 0.9105 | 0.9757 | 0.3700 | 0.5770 | 0.9153 | |
| 007 | 0.8711 | 0.9334 | 0.3874 | 0.5770 | 0.9230 | yes |
| 008 | 0.8492 | 0.9635 | 0.3913 | 0.6042 | 0.9214 | yes |
| 009 | 0.8121 | 0.9670 | 0.3643 | 0.5933 | 0.9184 | |
| 010 | 0.7898 | 0.9279 | 0.3981 | 0.6486 | 0.9125 | yes |
| 011 | 0.7691 | 0.9586 | 0.4068 | 0.6395 | 0.9127 | yes |
| 012 | 0.7429 | 0.9040 | 0.3952 | 0.6531 | 0.9207 | |
| 013 | 0.7352 | 0.9564 | 0.4126 | 0.6395 | 0.9207 | yes |
| 014 | 0.7003 | 0.9919 | 0.4135 | 0.6377 | 0.9217 | yes |
| 015 | 0.6929 | 0.9674 | 0.3961 | 0.6431 | 0.9150 | |
| 016 | 0.6724 | 0.9850 | 0.3787 | 0.6476 | 0.9035 | |
| 017 | 0.6476 | 1.0720 | 0.4135 | 0.6585 | 0.9130 | |
| 018 | 0.6353 | 1.0505 | 0.4222 | 0.6531 | 0.9219 | yes |
| 019 | 0.6073 | 0.9917 | 0.4261 | 0.6639 | 0.9125 | yes |
| 020 | 0.5884 | 1.0243 | 0.4116 | 0.6522 | 0.8888 | |
| 021 | 0.5707 | 1.0342 | 0.4319 | 0.6603 | 0.9156 | yes |
| 022 | 0.5523 | 1.0544 | 0.4386 | 0.6630 | 0.9105 | yes |
| 023 | 0.5315 | 1.0565 | 0.4744 | 0.6549 | 0.9238 | yes |
| 024 | 0.5118 | 1.0657 | 0.4773 | 0.6712 | 0.9265 | yes |
| 025 | 0.5001 | 1.0899 | 0.4870 | 0.6558 | 0.9231 | yes |
| 026 | 0.4880 | 1.0787 | 0.4966 | 0.6667 | 0.9229 | yes |
| 027 | 0.4830 | 1.0696 | 0.5043 | 0.6766 | 0.9207 | yes |
| 028 | 0.4712 | 1.0954 | 0.4976 | 0.6649 | 0.9269 | |
| 029 | 0.4649 | 1.0685 | 0.5024 | 0.6730 | 0.9253 | |
| 030 | 0.4581 | 1.0896 | 0.5043 | 0.6712 | 0.9264 |
Files
final/ (model.pt, config.json, label_map.json, test_metrics.json), best/, checkpoint_epoch_XXX/ (full resumable state),
results/ (history, plots, test predictions, audibility analysis), runs/ (TensorBoard), fxp2/ (model code), train.py.