Views
No views yet
dagger-phase1-proposed-librimix-3spk
warm-started from that checkpoint and continued on dataset.placement: scheduled scenes (see dagger.data.mixing.schedule_solo_then_overlap), which
guarantee every speaker both a solo enrollment window and participation in a
true depth-3 overlap in the same scene — something the Phase 1 chain-staggered
mixer cannot produce. Used for the Phase 2 depth-stratified experiment
comparing accumulation-free reconstruction (coarse_to_fine /
no_recursion) against the residual-deflation anti-pattern
(gated_deflation / ungated_deflation) that dagger's central claim
(CLAUDE.md §1) argues against.dagger-phase1-proposed-librimix-3spk (same architecture;
weights loaded via train.init_checkpoint, not random init).configs/phase2_librimix_3spk_train_scheduled.yaml
— same 2000-scene / 30-epoch scale as Phase 1 for comparability. Note: the
checkpoint actually hosted here is the epoch-25 mid-run save
(checkpoint_every: 5), not epoch 30 — the original training run hit a
Kaggle 12h execution ceiling during epoch 30; see the repo's CLAUDE.md
Phase 2 "Fine-tune attempt" note for the full story.configs/phase2_librimix_3spk_eval_finetuned.yaml.coarse_to_fine (0.21 dB) > gated_deflation (-2.19 dB) > ungated_deflation (-3.92 dB) at overlap depth 3, replicating the theoretically-predicted
order on an independently fine-tuned checkpoint. See the repo's CLAUDE.md
§5 Phase 2 notes and results/phase2_librimix_3spk_finetuned.csv for the
full table and honest caveats (the accumulation-specific gap did not widen
from fine-tuning alone — see the "did NOT sharpen" note in the same section).proposed_librimix_3spk_scheduled.pt — PyTorch state dict for the
fine-tuned extractor G. Same architecture/config shape as the Phase 1
checkpoint (cross_attn_blocks: 6, n_tokens: 8, etc. — must match, or
load_state_dict will raise).dagger repository and the Phase 1 checkpoint above.φ (NVIDIA TitaNet-Large,
nvidia/speakerverification_en_titanet_large) is CC-BY-4.0 licensed by
NVIDIA, loaded at runtime via NeMo, and is not bundled in this HF repo.
This repo's Apache-2.0 license covers only the extractor weights uploaded
here. See NOTICE
in the dagger repo for the canonical text.1from huggingface_hub import hf_hub_download
2import torch
3
4ckpt_path = hf_hub_download(
5 repo_id="AdityaAA2004/dagger-phase2-proposed-librimix-3spk-scheduled",
6 filename="proposed_librimix_3spk_scheduled.pt",
7)
8state_dict = torch.load(ckpt_path, map_location="cpu")