| --- |
| license: other |
| tags: [speech-emotion-recognition, wavlm, hubert, emotion2vec, arabic] |
| --- |
| |
| # TAFFC-2026-04-0506 — checkpoints |
|
|
| Companion checkpoints for *Adaptive WavLM-HuBERT Fusion Mechanism for Cross-Lingual |
| Speech Emotion Recognition: English-to-Arabic Transfer Learning* (IEEE TAFFC, major |
| revision). |
|
|
| **Start with [`MANIFEST.csv`](MANIFEST.csv)** — one row per checkpoint, saying in plain |
| English what model it is, which data it was trained on, which seed, what the file is |
| for, and whether it is worth keeping. 136 of 473 files are final models; 229 are |
| disposable intermediate state (183 GB). |
|
|
| ## Why the paths look like that |
|
|
| Paths mirror the run directory tree on the compute server. That is deliberate: the sync |
| tool decides "already uploaded, safe to delete locally" by comparing the repo path and |
| size against the local file, so renaming anything here would make every file look |
| un-uploaded and trigger a full re-upload of the repo. The meaning lives in |
| `MANIFEST.csv` instead of in the filenames. |
|
|
| ## Decoding a run directory name |
|
|
| `HPO_<regime>_<encoder>[_<corpus>][_Frozen]` |
|
|
| | piece | meaning | |
| |---|---| |
| | `Eng` | weights **initialised from** the English-trained checkpoint (transfer learning) | |
| | `Scr` | **from scratch** — SSL pre-training only, no English stage | |
| | `Fusion` / `5seeds` | the paper's gated fusion (WavLM+HuBERT, per-sample softmax gate) | |
| | `CrossAttn` | cross-attention fusion instead of the gate | |
| | `WavLM` / `HuBERT` | that encoder alone, no fusion | |
| | `E2V` | emotion2vec, a third unrelated encoder | |
| | `_BAVED` / `_EYASE` | trained on that corpus **alone**; absent = both pooled | |
| | `_Frozen` | encoder frozen, only a linear head trained | |
| | `LP_` | linear probe — dropped from the revision | |
|
|
| **The trap:** every `HPO_*` run trains *and* tests on **Arabic**. `Eng` describes the |
| initialisation, not the data. English-data runs have different names entirely |
| (`Combined`, `Cross_Attention`, `Emo2Vec_*`, and the `Arch_Ablation` variants). |
|
|
| ## File roles |
|
|
| | filename pattern | what it is | keep? | |
| |---|---|---| |
| | `FINAL_BEST_seed<N>.pth`, `best_model_seed<N>.pth` | the final model for that seed | **yes** | |
| | `wavlm_hubert_optimized_seed<N>.pth` | final model plus metadata | yes | |
| | `FINAL_BEST_epoch<N>_uar<X>.pth` | best-epoch snapshot during final training | maybe | |
| | `training_resume_seed<N>.pth` | mid-training optimiser state, resume only | no | |
| | `trial<N>_BEST_uar<X>.pth` | a single Optuna HPO trial | no | |
| | `epoch<N>_uar<X>.pth` | periodic snapshot, never read back | no | |
|
|
