DCASE 2026 Task 7 - Fine-Tuned Expert Aggregation

This repository contains the cleaned source code for our DCASE 2026 Challenge Task 7 submission on domain-agnostic incremental audio classification.

The submitted systems use full fine-tuned MCnn14 domain experts. The official D1 checkpoint is used only as the initialization source. A D2 expert is trained with device augmentation, a D3 expert is trained with gain-shift augmentation, and inference aggregates D2/D3 expert probabilities.

Submitted Systems

System	Inference	Test-time augmentation
S1	Entropy soft MoE, tau=3.0	No
S2	Entropy soft MoE, tau=4.0	No
S3	Entropy soft MoE, tau=3.0	Full-safe TTA
S4	Mean probability averaging	No

Repository Structure

.
├── README.md
├── requirements.txt
├── config_task7.py
├── domain_net.py
├── train_domain_aug_tta_routing_d1_router_fast.py
├── generate_fullfinetune_repro_artifacts.py
├── scripts/
│   └── run_task7_fullfinetune_repro.sh
└── docs/
    ├── method_summary.md
    ├── results_summary.md
    ├── results/
    └── report_tables/

Environment Setup

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

The code expects PyTorch with CUDA support for full training runs.

Required Data and Checkpoints

Datasets, evaluation audio, and model checkpoints are not included in this public repository. Obtain the official DCASE Task 7 resources separately.

Expected external paths can be supplied through environment variables:

export DATA_ROOT=/path/to/task7_data
export D1_CKPT=/path/to/checkpoint_D1.pth
export EVAL_ROOT=/path/to/evaluation/audio

DATA_ROOT should contain the official metadata and evaluation setup files, including evaluation_setup/development_train.txt and evaluation_setup/development_test.txt.

Training D2/D3 Experts

CUDA_VISIBLE_DEVICES=0 bash scripts/run_task7_fullfinetune_repro.sh

The runner trains plain MCnn14 D2/D3 experts from the official D1 checkpoint:

• D2 training augmentation: device augmentation
• D3 training augmentation: gain-shift augmentation
• Objective: cross entropy
• Optimizer: AdamW with cosine learning-rate schedule

Use SMOKE_TEST=1 for a one-epoch syntax and pipeline check without producing final submission artifacts:

SMOKE_TEST=1 CUDA_VISIBLE_DEVICES=0 bash scripts/run_task7_fullfinetune_repro.sh

Generating Submission Artifacts

After checkpoints exist, the runner calls generate_fullfinetune_repro_artifacts.py to compute development metrics, report tables, and submission-format outputs. Generated outputs are written under runs/, which is intentionally ignored by git.

Results

See docs/results_summary.md for the validation summary. Small CSV summaries are included under docs/results/ for reproducibility at the report level.

Notes

• Checkpoints (*.pth, *.pt, *.ckpt) are intentionally excluded.
• Official development/evaluation audio is intentionally excluded.
• Submission output CSV/ZIP files are intentionally excluded from this source-code repository.
• The official D1 checkpoint must be obtained from the DCASE Task 7 resources.

Citation / Task Link

Please refer to the official DCASE Challenge Task 7 page for task rules, data access, and evaluation protocol.

Contact

Semin Heo: smheo@seoultech.ac.kr