Update README: real FT checkpoint, link to training repo
Browse files
README.md
CHANGED
|
@@ -1,36 +1,13 @@
|
|
| 1 |
-
#
|
| 2 |
|
| 3 |
-
|
| 4 |
-
upstream noise-only MUSE checkpoint** (`g_best`, ~2.3 MB), used as a stand-in
|
| 5 |
-
for the real reverb-fine-tuned epoch-48 checkpoint until that file can be
|
| 6 |
-
retrieved from the iscohen_prj research cluster.
|
| 7 |
|
| 8 |
-
##
|
| 9 |
|
| 10 |
-
|
| 11 |
|
| 12 |
-
|
| 13 |
|
| 14 |
-
|
| 15 |
-
scp <cluster>:/rg/iscohen_prj/yairamr/code/Muse-Reverb-FN/checkpoints/all/epoch_48/g_00051852 \
|
| 16 |
-
./_workspace/release_artifacts/muse_reverb_e48.pt
|
| 17 |
-
```
|
| 18 |
|
| 19 |
-
|
| 20 |
-
(`HF_REPO_CKPT`, default `YairAmar/SE-Probe-models`). `scripts/setup.py`
|
| 21 |
-
already pulls from there, so notebook 06's inference cells will start
|
| 22 |
-
producing correct numbers automatically once the upload completes.
|
| 23 |
-
|
| 24 |
-
## Why this placeholder is safe
|
| 25 |
-
|
| 26 |
-
Notebook 06's published figure is computed from precomputed CKA results
|
| 27 |
-
(`results_demo/cka_reverb_muse_demo.parquet` or
|
| 28 |
-
`results_df/reverb/cka_reverb_muse_epoch_48.parquet`), not from running the
|
| 29 |
-
model — so the figure is correct regardless of which checkpoint is shipped.
|
| 30 |
-
Inference cells in notebook 06 are gated behind
|
| 31 |
-
`SE_PROBE_RUN_INFERENCE=1` and clearly labelled as placeholder-mode in the
|
| 32 |
-
notebook markdown, so a curious reader is warned before they get wrong
|
| 33 |
-
numbers.
|
| 34 |
-
|
| 35 |
-
See `_workspace/cluster_artifacts_needed.md` (item 1) for the broader
|
| 36 |
-
punch-list of cluster artefacts.
|
|
|
|
| 1 |
+
# SE-Probe model artefacts
|
| 2 |
|
| 3 |
+
Companion model checkpoints for [SE-Probe](https://github.com/YairAmar/SE-Probe).
|
|
|
|
|
|
|
|
|
|
| 4 |
|
| 5 |
+
## Files
|
| 6 |
|
| 7 |
+
- `muse_reverb_e48.pt` (2.3 MB): Reverb-fine-tuned MUSE checkpoint (epoch 48). PESQ 3.02 / STOI 0.944 on the 824x5 RIR test set; corresponds to the FT model probed in Section 4.2 of the paper. Trained from the upstream noise-only MUSE g_best, fine-tuned on RIR-Mega convolutive mixtures.
|
| 8 |
|
| 9 |
+
Training pipeline that produced this checkpoint: <https://github.com/YairAmar/muse-dereverb-ft>.
|
| 10 |
|
| 11 |
+
## License
|
|
|
|
|
|
|
|
|
|
| 12 |
|
| 13 |
+
MIT.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|