schmittzhuattencent/30

Spatial-Qwen

A spatial audio understanding framework that integrates spatial audio encoders with Qwen2.5-Omni LLM for spatial QA tasks.

Overview

This repository integrates:

• SELD233 Encoder: SELDNet-based spatial audio encoder (DCASE 2024 baseline, task 235/233)
• Spatial-BEATs Encoder: BEATs-based spatial audio encoder with local spatial encoding
• Qwen2.5-Omni: Multimodal LLM backbone
• Spatial Token Pipeline: FOA audio → spatial tokens → LLM injection via <|spatial|> placeholder

Architecture

FOA Audio [B, T, 4]
    ↓
[Spatial Encoder Choice]
├── SELD233: FeatureBridge → Backbone → SpatialAdapter → 2.5Hz tokens
└── BEATs:   SpatialBEATs (fused_spatial_embeddings) → Bridge → Projector
    ↓
Spatial Tokens [B, T_spat, D]  (2.5 Hz, 50 tokens per 20s)
    ↓
Projector → Qwen2.5-Omni LLM (injected at <|spatial|> positions)
    ↓
QA Answer

Repository Structure

Spatial-Qwen/
├── spatial_qwen/
│   ├── model/              # Qwen2.5-Omni spatial thinker, processor, configuration
│   ├── modules/            # Spatial adapter/bridge modules (seldnet233, SPUR)
│   ├── encoders/
│   │   ├── seldnet/        # DCASE SELDNet encoder (parameters, model, data)
│   │   └── beats/          # Spatial-BEATs encoder
│   ├── data/               # QA pair generation utilities
│   ├── utils/              # SELD metrics, spatial utility functions
│   └── ufb_banding/        # UFB filterbank framework
├── scripts/
│   ├── train_spatial_qa.py         # Main QA training script
│   ├── bench_spatial_qa.py         # Inference / benchmarking
│   ├── precompute_feature_cache.py # Pre-compute SELD features
│   ├── precompute_hidden_cache.py  # Pre-compute SELD hidden states
│   └── validate_spatial_pipeline.py
├── configs/
│   └── seld233_adapter_lora.yaml   # Training config
├── shell/
│   └── launch_train_seld233.sh     # Launch script
└── docs/
    ├── spatial_beats_design.md
    ├── baseline_experiments.md
    ├── seld233_implementation.md
    └── spatial_encoder_plan.md

Setup

pip install -r requirements.txt

Key Paths (configure via CLI args or config)

• Qwen2.5-Omni checkpoint: /path/to/Qwen2.5-Omni-7B
• SELD233 checkpoint: spatial_qwen/encoders/seldnet/ or external DCASE checkpoint
• SELD feature stats: pre-computed normalization stats directory
• QA dataset: prepared_datasets/starss23_foa_plus_29cls_20s/qa_pairs_v6c/

Training

# Stage 1: Train spatial adapter + projector only (frozen encoder)
torchrun --nproc_per_node=4 scripts/train_spatial_qa.py \
    --train-mode adapter_lora \
    --model-id /path/to/Qwen2.5-Omni-7B \
    --seld233-checkpoint /path/to/seld233.h5 \
    --qa-version v6c

# Stage 2: Full spatial + LoRA training
torchrun --nproc_per_node=4 scripts/train_spatial_qa.py \
    --train-mode spatial_lora \
    --resume-from /path/to/stage1_checkpoint

Spatial Token Baselines

See docs/baseline_experiments.md for IV-token and CNN-IV baseline designs.

BEATs Integration

See docs/spatial_beats_design.md for Spatial-BEATs → LLM bridge design.

Citation

Based on DCASE 2024 SELD baseline and Qwen2.5-Omni.