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 — 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 |