30 / docs /spatial_encoder_plan.md
schmittzhuattencent's picture
Add files using upload-large-folder tool
65ae8ca verified
|
Raw
History Blame Contribute Delete
13.3 kB
# SELDNet-233 Spatial Modality Integration Design
## Document Status
- Owner: Student Implementation
- Reviewer: Project Maintainer
- Target Repository: `/apdcephfs_cq10/share_1603164/user/schmittzhu/code/spur-qwen-2.5-omni`
- Related Baseline Repository: `/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline`
## 1. Background
This project aims to extend the current `Qwen2.5-Omni`-based system with a new **independent spatial modality** derived from a pretrained `SELDNet` encoder.
The current system already supports:
- text
- image
- video
The current SPUR-related spatial path in the Omni codebase is not a truly independent modality. Instead, it computes spatial features and fuses them back into the audio encoder branch. This is
**not** the desired design for this project.
- keep the **original audio encoder unchanged**
- use the same FOA input audio in **two parallel paths**
- inject a new set of **spatial tokens** into the LLM as a separate modality
## 2. Objective
Given a 4-channel FOA audio input:
1. The **W channel** should continue to go through the original Omni mono audio pipeline.
2. The full **4-channel FOA** should go through a pretrained `SELDNet-233` encoder.
3. The `SELDNet-233` encoder should provide a sequence of **spatial tokens** derived from its **MHSA output**, not from its final localization/classification heads.
4. These spatial tokens should be projected into the LLM hidden space and injected into the thinker decoder input as a new modality `<|spatial|>`.
## 3. Non-Goals
- modify the existing SELD baseline source code
- modify existing SELD training or inference behavior
- use SELDNet final `accdoa` or `sed_logits` as the LLM input representation
## 4. Fixed Dependencies
- Path: `/apdcephfs_cq10/share_1603164/user/schmittzhu/code/DCASE2024_seld_baseline`
### 4.2 SELD checkpoint to use
### 4.3 SELD task configuration
- Task ID: `233`
### 4.4 SELD model definition
## 5. Key Design Decision
### 5.1 Spatial representation source
The spatial representation must come from the **shared temporal representation after the final MHSA block** in `SELDNet`, not from:
- final `accdoa`
- final `sed_logits`
- explicit detection heads
### 5.2 Why this representation
This tensor is:
- temporally ordered
- dense
- scene-level
- richer than detection output heads
- better suited for LLM conditioning
## 6. Target Architecture
### 6.1 Input
### 6.2 Dual-path processing
- take FOA `W channel`
- convert to mono
- feed into original Omni audio encoder
- keep existing tokenization rate unchanged
- expected token rate: about `25 Hz`
Path B: New spatial path
- keep all 4 FOA channels
- convert to SELD baseline-compatible input features
- feed into pretrained `SELDNet-233`
- produce spatial token sequence
- project into LLM hidden size
- inject as `<|spatial|>` modality
### 6.3 Final LLM view
- normal text tokens
- normal audio tokens from original audio encoder
- optional image/video tokens
- new spatial tokens from SELDNet-233 branch
## 7. Temporal Resolution Requirements
### 7.1 Audio token frequency
Keep current Omni audio token frequency unchanged:
- approximately `25 Hz`
### 7.2 Spatial token frequency
Spatial tokens should be lower frequency:
- target `2.5 Hz`
This is intentional because:
- lower token rate reduces prompt length
- lower token rate is sufficient for spatial reasoning
Recommended first version:
- `h_seld`: `[B, T_seld, 128]`
- `spatial_tokens`: `[B, T_spat, 256]`
- `projected_spatial`: `[B, T_spat, D_llm]`
Where:
- `T_spat = 5 * seconds`
## 8. Data Flow
1. Input FOA waveform enters processor.
2. Processor keeps original audio path intact:
- use `W channel` for original audio encoder.
3. Processor also passes full FOA waveform as `spatial_audio`.
- `spatial_audio` for new spatial path
5. New SELD spatial adapter computes:
- SELD MHSA output
- `2.5 Hz` spatial tokens
6. Spatial projector maps tokens to LLM hidden dimension.
8. Thinker injects projected spatial embeddings into `inputs_embeds`.
9. LLM decoder consumes them as a true multimodal prefix.
## 9. Required Code Work
### A. `qwen2_5_omni_spur/modules/seldnet233_backbone.py`
Purpose:
- build baseline `SeldModel` with task `233`
- load pretrained checkpoint
- freeze model by default
- register a forward hook on the final MHSA output
Expected behavior:
- input: baseline-compatible SELD features
- output: `h_seld [B, T_seld, D_seld]`
- do not modify baseline code
- use import-time wrapping or dynamic module loading
- use forward hook on the last MHSA normalization output
### B. `qwen2_5_omni_spur/modules/seldnet233_feature_bridge.py`
Purpose:
- convert FOA waveform into the feature representation expected by SELD task `233`
Expected behavior:
- output: baseline feature tensor `[B, C, T_feat, F_feat]`
Implementation notes:
- must match SELD task `233` feature config exactly
- sample rate must be `16k`
- should reuse baseline feature logic where possible
Purpose:
- build a full SELD-based spatial token extractor
Internal structure:
- feature bridge
- SELD backbone
- spatial token head
Expected behavior:
- output:
- `spatial_tokens [B, T_spat, D_spat]`
Implementation notes:
- token rate target must be `2.5 Hz`
- default token dim should be `256`
- token head may use:
- pooling
- MLP
- recommended first implementation:
### D. Optional: `qwen2_5_omni_spur/modules/spatial_token_projector.py`
Purpose:
- project SELD spatial tokens into LLM hidden dimension
## 9.2 Existing files to modify
### A. `qwen2_5_omni_spur/configuration_qwen2_5_omni.py`
Add config fields:
- `spatial_token_index`
- `use_seld233_spatial_modality`
- `seld233_checkpoint_path`
- `seld233_token_dim`
- `seld233_token_rate_hz`
- `seld233_projector_hidden_dim`
- optional `seld233_freeze_backbone`
- `attribute_map` to include `spatial_token_id`
### B. `qwen2_5_omni_spur/modules/__init__.py`
Export new modules:
- `SeldNet233Backbone`
- optional new projector class
### C. `qwen2_5_omni_spur/processing_qwen2_5_omni.py`
Add a new special token:
- `<|spatial|>`
- store `self.spatial_token`
- optionally accept `spatial_tokens` and `spatial_token_lengths`
- expand `<|spatial|>` into repeated placeholders based on spatial token length
- expose spatial fields through `model_input_names`
Important constraint:
- original W-channel mono audio path must remain unchanged
### D. `qwen2_5_omni_spur/modeling_qwen2_5_omni.py`
Required changes:
- instantiate SELD spatial adapter in thinker
- accept `spatial_audio` in thinker forward
- compute spatial embeddings
- inject them via `masked_scatter`
- update multimodal rope logic
- route spatial fields through generation
### E. `training-qwen-omni/train_spur_spatial_only_hf.py`
Update collator:
- additionally pass full FOA as `spatial_audio`
### F. `qwen-omni-inference.py`
Update inference script:
- register `<|spatial|>` token
- construct prompt with spatial placeholder
- validate both paths run together
## 10. Processor Design
### 10.1 Prompt convention
- `<|spatial|>`
Example:
- `<|AUDIO|><|spatial|> Please answer the question using both sound content and spatial cues.`
### 10.2 Placeholder expansion behavior
- `T_spat` repeated spatial token positions
This mirrors current audio/image/video placeholder logic.
### 10.3 Recommended input mode
- processor receives `spatial_audio`
Fallback mode:
- processor receives precomputed `spatial_tokens`
## 11. SELD Path Design
### 11.1 Backbone loading
The SELD branch must:
- use task `233`
- load the fixed checkpoint above
- stay frozen in the first training stage
### 11.2 Feature extraction
The spatial branch must operate on:
- full 4-channel FOA
- `16kHz`
- baseline-compatible feature format
### 11.3 Hidden representation extraction
The representation to extract is:
- final MHSA output
- not final heads
- not sed logits
### 11.4 Tokenization
The spatial token head must:
- preserve temporal ordering
- reduce token rate to `2.5 Hz`
- output a fixed hidden dim, recommended `256`
### 12.1 New thinker members
Add:
- optional `self.seld233_spatial_norm`
### 12.2 Forward inputs
Thinker forward should support:
- `spatial_audio`
- `spatial_tokens`
- `spatial_token_lengths`
Priority:
- if `spatial_tokens` provided, use them directly
### 12.3 Injection point
Injection should happen in the same stage where current code merges:
- text
- audio
- video
The new logic should:
1. compute `projected_spatial`
2. build `spatial_mask = (input_ids == spatial_token_id)`
3. inject with `masked_scatter`
Before scatter, verify:
If not, raise an error.
### 13.1 Why needed
Adding a new modality requires updating multimodal position ID construction.
### 13.2 Required additions
- add `spatial_token_id`
- add `spatial_seqlens`
- count spatial modality occurrences
- add `remain_spatials`
- add spatial branch when selecting next multimodal segment
### 13.3 Spatial position assignment
Spatial token positions should be assigned:
- length equal to `spatial_token_lengths`
- consistent with `2.5 Hz` token rate
## 14. Generation Path Changes
Top-level `generate()` must forward:
- `spatial_audio`
- `spatial_tokens`
- `spatial_token_lengths`
to thinker generation.
### 14.2 Talker interaction
Recommended first version:
- spatial positions should be zeroed before passing thinker hidden states into talker
Reason:
- spatial modality is meant for reasoning
- not necessarily for speech token generation conditioning
## 15. Training Strategy
### 15.1 Stage 1
Freeze:
- SELD backbone
Train only:
- spatial token head
- spatial projector
- LoRA / lightweight LLM adapters if used
- last spatial token head layers
- projector
- maybe a small part of SELD adapter
### 15.3 Stage 3
Optional advanced finetuning:
- limited SELD backbone finetuning
- only after full pipeline is stable
## 16. Validation Plan
### 16.1 Unit-level validation
- checkpoint loads successfully
- MHSA output is captured successfully
- spatial token head outputs expected shape
- token rate is `2.5 Hz`
### 16.2 Processor validation
Validate:
- spatial fields appear in batch output
- original audio path is unchanged
Validate:
- spatial projector output matches LLM hidden size
- no length mismatch
- no dtype/device mismatch
### 16.4 Rope validation
Validate:
- generation does not crash when `<|spatial|>` is present
Validate:
- original audio tokens still exist
- spatial tokens are injected independently
- text generation completes
## 17. Risks
### 17.1 Feature mismatch risk
If SELD features differ from training-time `233` config, checkpoint utility may collapse.
Mitigation:
- strictly reuse baseline feature logic
If hook location is incorrect, extracted representation may not correspond to post-MHSA scene representation.
Mitigation:
- verify against baseline layer order
- inspect shape and stability
### 17.3 Rope mismatch
If `<|spatial|>` positions are not included in multimodal rope indexing, model behavior may degrade or fail.
Mitigation:
- treat spatial exactly as a fourth modality in rope construction
### 17.4 Prompt length growth
Adding spatial tokens increases context length.
Mitigation:
## 18. Final Requirements Checklist
- [ ] Use checkpoint `233_merged29_foa_16k_sedwarmup_v1_dev_split0_multiaccdoa_foa_best_full_model.h5`
- [ ] Keep original audio encoder unchanged
- [ ] Audio path uses `W channel` only
- [ ] Spatial path uses all 4 FOA channels
- [ ] Spatial representation comes from final SELD MHSA output
- [ ] Spatial token rate is `2.5 Hz`
- [ ] Spatial token dimension defaults to `256`
- [ ] Spatial tokens are injected as a true independent modality
- [ ] `<|spatial|>` placeholder is supported in processor
- [ ] RoPE logic includes spatial modality
- [ ] End-to-end generate path runs successfully
## 19. Recommended Implementation Order
1. Implement `seldnet233_backbone.py`
2. Implement `seldnet233_feature_bridge.py`
3. Implement `seldnet233_spatial_adapter.py`
4. Add config fields
5. Add processor `<|spatial|>` support
6. Add thinker spatial encoder + projector
7. Add spatial `masked_scatter`
8. Update `get_rope_index()`
9. Update `generate()`
10. Update collator and inference script
11. Run shape-only validation
12. Run end-to-end generation validation
## 20. Reviewer Notes
This design intentionally separates:
- content understanding from the original audio encoder
- spatial reasoning from the SELDNet-based encoder
The design is successful only if the final model sees:
- mono audio content tokens at normal resolution
- spatial tokens at low resolution
- both as independent modalities inside the thinker input stream