docs/CAUTION.md

⚠️ CAUTION: WindOneSound V3.2 Architecture & Training Guide

1. Executive Summary

WindOneSound is a Multimodal Reasoning Engine. It listens to audio, builds a causal graph of events, remembers context, and generates natural language. It is NOT a simple transcriber.

2. The Data Flow Pipeline

Stage A: Perception (Frozen)

• Input: Raw Audio → BEATs → [B, T, 768].
• Module: PrimitiveExtractor.
• Output: Dense Primitives [B, T, 512].
• Status: FROZEN. Load weights from primitive_extractor_esc50_best.pt.

Stage B: Reasoning Core (Trainable)

1. SourceMemory: Biases features by speaker/source identity.
2. TimeConditioning: Injects sinusoidal time embeddings. CRITICAL: Added INSIDE CausalGraph, not LLM.
3. CausalGraphNetwork: Builds entity-event relationships using bidirectional attention. Uses Time Embeddings to order events.
4. COGENMemory: Long-term storage with Top-K slot addressing and Decay Factor (0.995).

• Output: read_out [B, T, 512] (Fully contextualized).

Stage C: Hybrid Discretization (The Bridge)

• Module: VQPrimitiveTokenizer.
• Outputs:
  1. z_q: Quantized continuous vectors [B, T, 512].
  2. tokens: Discrete integer IDs [B, T] (0-1023).
• Purpose: Filters noise and creates a stable "Audio Vocabulary".

Stage D: Generation Heads (Trainable)

• LLM Decoder: Generates text.
  • Context Input: HYBRID VECTOR (See Section 3).
  • Text Input: Question/Prompt tokens.
• Voice Head: Synthesizes Mel-Spectrograms from RAW read_out (no quantization artifacts).
• Agent Head: Outputs action vectors from pooled read_out.

3. The Hybrid Context Strategy (VQ + Projection)

To maximize performance, we do NOT choose between VQ Tokens and Continuous Projections. We USE BOTH.

Formula: Final_Context[t] = Linear_Projection(z_q[t]) + Learned_Embedding(VQ_Token_ID[t])

Why?

• Linear Projection (z_q): Carries rich semantic detail (timbre, pitch, emotion). Essential for descriptive accuracy.
• Learned Embedding (Token ID): Carries structural information (event boundaries, repetition patterns). Helps the LLM align text timestamps with audio events.

Implementation Note: In output_heads.py, ensure DualOutputHeads.forward constructs this hybrid tensor before passing it to self.text_decoder.

4. Training Strategy (Strict Phasing)

Phase 1: Validation (Done ✅)

• Verify Primitive Extractor quality (>90% ESC-50 Acc).

Phase 2: Codebook Stabilization (Done ✅)

• Train VQ Tokenizer on frozen primitives. Goal: >95% code usage.

Phase 3: Reasoning Warm-up

• Trainable: SourceMemory, CausalGraph, COGENMemory.
• Frozen: PE, VQ, LLM.
• Task: Contrastive Learning or Classification to teach temporal/causal structure.

Phase 4: Joint Multimodal Training

• Trainable: Everything EXCEPT PrimitiveExtractor.
• Optimizer: AdamW with Parameter Groups:
  • VQ Codebook: LR=1e-5
  • Graph/Memory: LR=1e-4
  • LLM Decoder: LR=5e-4
  • Heads: LR=1e-3
• Loss: L_text + 0.1*L_vq + 0.01*L_mel

5. How Time Conditioning Works

Time embeddings are added to primitive features BEFORE the Causal Graph processes them. This allows the Graph's attention mechanism to distinguish between "Event A at t=0" and "Event B at t=5", enabling causal ordering (Cause → Effect). The LLM sees the result of this ordering, not the raw time signals.

6. QA Capability

The model answers questions via Conditional Generation:

1. User asks: "Was there a dog?"
2. Text Encoder processes question.
3. Cross-Attention looks at Hybrid Audio Context.
4. If "Dog" concepts are present in z_q and structured correctly by Graph, LLM predicts "Yes".

7. Common Pitfalls

• NaN Loss: Clip gradients (max_norm=1.0). Lower VQ LR.
• Hallucinations: Ensure src_key_padding_mask is used so LLM ignores silent audio frames.
• Memory Overflow: Monitor COGEN novelty scores; adjust decay if slots saturate.