From 40d29bcaf86cea4eccfaeb7770251ce644a14737 Mon Sep 17 00:00:00 2001
From: Ethanfel <ethan.fel@ts-pc.fr>
Date: Fri, 10 Apr 2026 16:48:21 +0200
Subject: [PATCH] feat: add experiment configs for logit+cosine combo and
 BigVGAN decoder fine-tuning

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 experiments/bigvgan_optimized_dataset.json | 35 ++++++++++++++++++++++
 experiments/lora_logit_cosine_combo.json   | 33 ++++++++++++++++++++
 2 files changed, 68 insertions(+)
 create mode 100644 experiments/bigvgan_optimized_dataset.json
 create mode 100644 experiments/lora_logit_cosine_combo.json

diff --git a/experiments/bigvgan_optimized_dataset.json b/experiments/bigvgan_optimized_dataset.json
new file mode 100644
index 0000000..d8ecc31
--- /dev/null
+++ b/experiments/bigvgan_optimized_dataset.json
@@ -0,0 +1,35 @@
+{
+  "name": "bigvgan_optimized_dataset",
+  "description": "BigVGAN fine-tuning on optimized dataset (134 clips, 44.1kHz, LUFS-normalized). Standard mode (no LoRA) — trains decoder to faithfully reconstruct target domain audio from mel spectrograms. Uses optimal config from prior sweeps.",
+  "data_dir": "/media/unraid/davinci/Selva/BJ/features_v2_improved/",
+  "output_root": "/media/unraid/davinci/Selva/BJ/experiment/bigvgan_optimized_dataset",
+  "base": {
+    "train_mode": "snake_alpha_only",
+    "steps": 5000,
+    "lr": 1e-4,
+    "batch_size": 8,
+    "segment_seconds": 0.5,
+    "lambda_l2sp": 1e-3,
+    "use_gafilter": true,
+    "gafilter_kernel_size": 9,
+    "lambda_phase": 1.0,
+    "save_every": 1000,
+    "seed": 42
+  },
+  "experiments": [
+    {
+      "id": "standard_5k",
+      "description": "Standard mode: mel from clean FLAC → BigVGAN → reconstruct FLAC. No LoRA. Directly improves VAE roundtrip quality."
+    },
+    {
+      "id": "disc_fm_5k",
+      "description": "Standard mode + discriminator feature matching. Tests if perceptual loss helps on clean audio reconstruction.",
+      "discriminator_path": "/media/unraid/davinci/Selva/BJ/experiment/bigvgan_discriminator_optimizer.pt"
+    },
+    {
+      "id": "standard_10k",
+      "description": "Extended 10k steps. More data passes on 134 clips may extract more from the optimized dataset.",
+      "steps": 10000
+    }
+  ]
+}
diff --git a/experiments/lora_logit_cosine_combo.json b/experiments/lora_logit_cosine_combo.json
new file mode 100644
index 0000000..35dc52e
--- /dev/null
+++ b/experiments/lora_logit_cosine_combo.json
@@ -0,0 +1,33 @@
+{
+  "name": "lora_logit_cosine_combo",
+  "description": "Combine the two best findings from optimized dataset sweep: logit-normal timestep sampling + cosine LR schedule. Both individually outperformed baseline by large margins (56% and 68% lower loss). Tests if gains stack.",
+  "data_dir": "/media/unraid/davinci/Selva/BJ/features_v2_improved/",
+  "output_root": "/media/unraid/davinci/Selva/BJ/experiment/lora_logit_cosine_combo",
+  "base": {
+    "rank": 128,
+    "lr": 3e-4,
+    "steps": 5000,
+    "batch_size": 16,
+    "warmup_steps": 100,
+    "save_every": 1000,
+    "seed": 42,
+    "init_mode": "pissa",
+    "use_rslora": true,
+    "target": "attn.qkv",
+    "timestep_mode": "uniform",
+    "lr_schedule": "constant"
+  },
+  "experiments": [
+    {
+      "id": "logit_normal_cosine",
+      "description": "Logit-normal timesteps + cosine LR decay. Combining the two best individual improvements.",
+      "timestep_mode": "logit_normal",
+      "lr_schedule": "cosine"
+    },
+    {
+      "id": "logit_normal_control",
+      "description": "Control: logit-normal only (constant LR). Reproduces previous winner for direct comparison.",
+      "timestep_mode": "logit_normal"
+    }
+  ]
+}