File size: 10,605 Bytes
ebcf321
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
"""
Whisper encoder with HGA-modulated self-attention.

================================================================================
V1 (2A): HGA now includes a Möbius bias term that breaks exp/log cancellation,
making the curvature c a real, gradient-bearing parameter.

Modulation formula (per Q/K/V projection, per layer):

    W_HGA = log_0^c( ( diag(s) ⊗_c exp_0^c(W_ref) ) ⊕_c exp_0^c(b) )

Without b, the chain reduces to s ⊙ W_ref (DoRA-like, c does nothing — this
is the SER paper's formula). With b ≠ 0, the Möbius addition step entangles c
with the norms of q and b in a way that cannot be algebraically cancelled.

Key properties:
  - b tiny-random-init (std=1e-4) → init is numerically ≈ s ⊙ W_ref but with
    b ≠ 0, so c receives gradient signal from step 0. (Zero-init would freeze
    c at a saddle ∂L/∂c = 0.)
  - All layers use the same (c_min, c_init, c_max) bounds; layer-aware bucketing
    removed because b makes c a real parameter that learns its own per-layer
    optimum without artificial floors.
================================================================================
"""
import math
import logging
from typing import List, Dict, Any, Optional

import torch
import torch.nn as nn
import torch.nn.functional as F

from .hyperbolic_ops import (
    exp_map_zero, log_map_zero, mobius_add, clamp_norm,
    LearnableCurvature,
)

logger = logging.getLogger(__name__)


class HGALinear(nn.Module):
    """Drop-in replacement for nn.Linear that applies HGA weight modulation
    with a Möbius bias term.

    Stores a frozen reference weight W_ref. At forward time:
        p     = exp_0^c(W_ref)                                  # rows → ball
        v     = log_0^c(p)                                      # = W_ref (id)
        q     = exp_0^c(diag(s) · v)        = exp_0^c(s ⊙ W_ref)# Möbius scale
        b_pt  = exp_0^c(b)                                      # bias → ball
        r     = q ⊕_c b_pt                  # Möbius add — c becomes essential
        W_mod = log_0^c(r)                                      # ball → tangent
        output = x @ W_mod^T + bias_orig

    Trainable: s (d_in,), b (d_in,), c (via curvature_module)
    Frozen:    W_ref, bias_orig (from pretrained Whisper)
    """

    def __init__(self, original_linear: nn.Linear,
                 s: nn.Parameter, b: nn.Parameter,
                 curvature_module: nn.Module):
        super().__init__()
        # Frozen reference weight (rows are the d_out "row vectors" in R^{d_in})
        self.register_buffer("W_ref", original_linear.weight.data.clone().float())
        # Keep original bias (frozen but used in forward)
        if original_linear.bias is not None:
            self.register_buffer("bias", original_linear.bias.data.clone())
        else:
            self.bias = None
        # Learnable HGA parameters (shared from the per-layer container)
        self.s = s
        self.b = b
        self.curvature = curvature_module

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        c = self.curvature().float()

        # Step 1: rows of W_ref → Poincaré ball
        p = exp_map_zero(self.W_ref, c)                # (d_out, d_in)
        # Step 2: Möbius diagonal scaling diag(s) ⊗_c p
        #   = exp_0^c( s ⊙ log_0^c(p) ) = exp_0^c( s ⊙ W_ref )
        v = log_map_zero(p, c)                         # = W_ref (cancellation)
        v_scaled = v * self.s.float().unsqueeze(0)
        q = exp_map_zero(v_scaled, c)                  # (d_out, d_in) in ball
        # Step 3: Möbius bias addition — broadcasts b across d_out rows
        b_pt = exp_map_zero(self.b.float(), c)         # (d_in,) in ball
        b_pt_b = b_pt.unsqueeze(0).expand_as(q)        # (d_out, d_in)
        r = mobius_add(q, b_pt_b, c)                   # (d_out, d_in)
        r = clamp_norm(r, c)                           # numerical safety
        # Step 4: log_map back to tangent → modulated weight
        W_mod = log_map_zero(r, c)                     # (d_out, d_in)

        with torch.amp.autocast("cuda", enabled=False):
            return F.linear(x.float(), W_mod.float(),
                            self.bias.float() if self.bias is not None else None).to(x.dtype)


class HGAWhisperEncoder(nn.Module):
    """Whisper encoder with HGA-modulated Q/K/V on all 32 layers.

    Architecture:
      1. Load Whisper encoder, freeze every original parameter.
      2. For each layer, create one shared LearnableCurvature c^(l) plus three
         pairs of (s, b) — for q_proj, k_proj, v_proj.
      3. Replace q_proj/k_proj/v_proj with HGALinear wrappers that compute
         the modulated weight on the fly.
      4. Register forward hooks to capture multi-scale features for EMCA.

    Trainable params per layer:
        3 × d_model (s_Q, s_K, s_V)  +  3 × d_model (b_Q, b_K, b_V)  +  1 (c)
      = 6 × d_model + 1
    For d_model=1280 → 7,681 per layer × 32 layers ≈ 246K total (HGA only).
    """

    output_dim = 1280
    output_frame_rate_hz = 50.0

    def __init__(self, model_path: str, extract_layers: List[int],
                 num_encoder_layers: int = 32,
                 hga_c_init: float = 1.0,
                 hga_c_min: float = 0.001,
                 hga_c_max: float = 8.0,
                 hga_b_init_std: float = 1.0e-4):
        super().__init__()
        self.extract_layers = sorted(extract_layers)
        self.num_encoder_layers = num_encoder_layers
        self.hga_c_init = hga_c_init
        self.hga_c_min = hga_c_min
        self.hga_c_max = hga_c_max
        self.hga_b_init_std = hga_b_init_std

        # --- Load Whisper encoder ---
        from transformers import WhisperModel
        whisper = WhisperModel.from_pretrained(model_path)
        self.encoder = whisper.encoder
        del whisper

        # Freeze ALL original encoder parameters
        for p in self.encoder.parameters():
            p.requires_grad = False

        # --- Create HGA params and inject into Whisper ---
        self.hga_layers = nn.ModuleList()
        d = self.output_dim
        for i, layer in enumerate(self.encoder.layers):
            attn = layer.self_attn

            # Shared curvature for Q/K/V of this layer
            curvature = LearnableCurvature(
                init_value=hga_c_init, c_min=hga_c_min, c_max=hga_c_max
            )

            # Diagonal scaling: identity (s=1) at init → first-step output ≈ W_ref
            s_q = nn.Parameter(torch.ones(d))
            s_k = nn.Parameter(torch.ones(d))
            s_v = nn.Parameter(torch.ones(d))

            # Möbius bias: tiny random init so b ≠ 0 from step 0 and ∂L/∂c ≠ 0
            b_q = nn.Parameter(torch.randn(d) * hga_b_init_std)
            b_k = nn.Parameter(torch.randn(d) * hga_b_init_std)
            b_v = nn.Parameter(torch.randn(d) * hga_b_init_std)

            # Replace q/k/v_proj with HGA-modulated versions
            attn.q_proj = HGALinear(attn.q_proj, s_q, b_q, curvature)
            attn.k_proj = HGALinear(attn.k_proj, s_k, b_k, curvature)
            attn.v_proj = HGALinear(attn.v_proj, s_v, b_v, curvature)

            # Container so optimizer sees these params
            cont = nn.Module()
            cont.curvature = curvature
            cont.s_q, cont.s_k, cont.s_v = s_q, s_k, s_v
            cont.b_q, cont.b_k, cont.b_v = b_q, b_k, b_v
            self.hga_layers.append(cont)

        logger.info(
            f"Whisper encoder: {num_encoder_layers} layers, "
            f"all Q/K/V wrapped in HGALinear "
            f"(c_init={hga_c_init}, c_min={hga_c_min}, c_max={hga_c_max}, "
            f"b_std={hga_b_init_std})"
        )

        # --- Feature capture hooks ---
        self._features: Dict[int, torch.Tensor] = {}
        self._hooks = []
        for idx, layer in enumerate(self.encoder.layers):
            if idx in self.extract_layers:
                self._hooks.append(
                    layer.register_forward_hook(self._make_hook(idx))
                )

    def _make_hook(self, layer_idx: int):
        def hook_fn(module, input, output):
            self._features[layer_idx] = (
                output[0] if isinstance(output, tuple) else output
            )
        return hook_fn

    def forward(self, mel_input: torch.Tensor) -> List[torch.Tensor]:
        """Run Whisper with HGA-modulated attention.

        Args:
            mel_input: (B, n_mels, T_mel)
        Returns:
            List of (B, T, 1280) tensors, one per extract_layer (sorted).
        """
        encoder_dtype = self.encoder.layer_norm.weight.dtype
        mel = mel_input.to(dtype=encoder_dtype)

        self._features.clear()
        _ = self.encoder(mel)

        features = []
        for ln in self.extract_layers:
            if ln not in self._features:
                raise RuntimeError(
                    f"Layer {ln} not captured. Got: {sorted(self._features.keys())}"
                )
            features.append(self._features[ln])
        return features

    def num_audio_frames(self, audio_samples_16khz: int) -> int:
        return min(math.ceil(audio_samples_16khz / 320), 1500)

    # ---- Diagnostics ----

    def get_hga_diagnostics(self) -> Dict[str, float]:
        """Per-layer scalars logged to TensorBoard."""
        diag = {}
        for i, hga in enumerate(self.hga_layers):
            diag[f"hga/c_L{i}"] = hga.curvature().item()
            diag[f"hga/s_q_mean_L{i}"] = hga.s_q.data.mean().item()
            diag[f"hga/s_v_mean_L{i}"] = hga.s_v.data.mean().item()
            # b norms — key indicator: if these stay ~0, c won't learn
            diag[f"hga/b_q_norm_L{i}"] = hga.b_q.data.norm().item()
            diag[f"hga/b_k_norm_L{i}"] = hga.b_k.data.norm().item()
            diag[f"hga/b_v_norm_L{i}"] = hga.b_v.data.norm().item()
        return diag

    def get_curvatures_summary(self) -> str:
        """Compact c-per-layer string for log lines (grouped 8 per row)."""
        vals = [f"{hga.curvature().item():.4f}" for hga in self.hga_layers]
        groups = []
        for i in range(0, len(vals), 8):
            groups.append("/".join(vals[i:i+8]))
        return " | ".join(groups)

    def get_b_norm_summary(self) -> str:
        """Compact b_q-per-layer norm string (grouped 8 per row).
        Used to verify the 'b grows → c learns' training dynamic."""
        vals = [f"{hga.b_q.data.norm().item():.4f}" for hga in self.hga_layers]
        groups = []
        for i in range(0, len(vals), 8):
            groups.append("/".join(vals[i:i+8]))
        return " | ".join(groups)