models/audio_model.py

import torch
import torch.nn as nn
from transformers import Wav2Vec2Model


# ── Shallow frozen wav2vec2 extractor (from train_val.ipynb) ─────────────────

class FrozenExtractor(nn.Module):
    """
    Loads facebook/wav2vec2-base, keeps only the CNN feature extractor
    + the first n_shallow transformer layers, freezes everything.
    Input : (B, T_wave)  raw float32 waveform, normalised to [-1, 1]
    Output: (B, T', 768) contextual features
    """

    def __init__(self, model_name: str, n_shallow: int):
        super().__init__()
        _w2v = Wav2Vec2Model.from_pretrained(model_name)
        self.feat_extractor  = _w2v.feature_extractor
        self.feat_projection = _w2v.feature_projection
        self.pos_conv_embed  = _w2v.encoder.pos_conv_embed
        self.layer_norm      = _w2v.encoder.layer_norm
        self.enc_dropout     = _w2v.encoder.dropout
        self.shallow_layers  = nn.ModuleList(_w2v.encoder.layers[:n_shallow])
        del _w2v
        torch.cuda.empty_cache()
        for p in self.parameters():
            p.requires_grad = False

    @torch.no_grad()
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        h = self.feat_extractor(x)
        h = h.transpose(1, 2)
        h, _ = self.feat_projection(h)
        pe = self.pos_conv_embed(h)
        h  = self.layer_norm(h + pe)
        h  = self.enc_dropout(h)
        for layer in self.shallow_layers:
            h = layer(h)[0]
        return h  # (B, T', 768)


# ── Deep audio inference modules (from test__1_.ipynb) ───────────────────────

class LinearProjection(nn.Module):
    def __init__(self, in_dim: int, out_dim: int, dropout: float):
        super().__init__()
        self.proj = nn.Sequential(
            nn.Linear(in_dim, out_dim),
            nn.LayerNorm(out_dim),
            nn.GELU(),
            nn.Dropout(dropout),
        )

    def forward(self, x: torch.Tensor, mask: torch.Tensor) -> torch.Tensor:
        return self.proj(x) * mask.unsqueeze(-1).float()


class AGRU(nn.Module):
    def __init__(self, dim: int, num_heads: int):
        super().__init__()
        self.gru  = nn.GRU(dim, dim // 2, batch_first=True, bidirectional=True)
        self.attn = nn.MultiheadAttention(embed_dim=dim, num_heads=num_heads, batch_first=True)
        self.norm = nn.LayerNorm(dim)

    def forward(self, x: torch.Tensor, mask: torch.Tensor) -> torch.Tensor:
        gru_out, _  = self.gru(x)
        attn_out, _ = self.attn(gru_out, gru_out, gru_out, key_padding_mask=~mask)
        out = self.norm(gru_out + attn_out) * mask.unsqueeze(-1).float()
        return torch.cat([gru_out, out], dim=-1)  # (B, T, dim*2)


class AttentionPooling(nn.Module):
    def __init__(self, dim: int):
        super().__init__()
        self.attn = nn.Linear(dim, 1)

    def forward(self, x: torch.Tensor, mask: torch.Tensor) -> torch.Tensor:
        scores  = self.attn(x).squeeze(-1).masked_fill(~mask, float("-inf"))
        weights = torch.softmax(scores, dim=1)
        return (x * weights.unsqueeze(-1)).sum(dim=1)


class PersonalityRegressor(nn.Module):
    def __init__(
        self,
        in_dim: int,
        ffn_dim: int,
        mlp_hidden: int,
        num_traits: int,
        num_heads: int,
        dropout: float,
    ):
        super().__init__()
        enc_layer = nn.TransformerEncoderLayer(
            d_model=in_dim,
            nhead=num_heads,
            dim_feedforward=ffn_dim,
            dropout=dropout,
            batch_first=True,
        )
        self.transformer = nn.TransformerEncoder(enc_layer, num_layers=1)
        self.pool        = AttentionPooling(in_dim)
        self.head        = nn.Sequential(
            nn.LayerNorm(in_dim),
            nn.Linear(in_dim, mlp_hidden),
            nn.GELU(),
            nn.Dropout(0.3),
            nn.Linear(mlp_hidden, num_traits),
            nn.Sigmoid(),
        )

    def forward(self, x: torch.Tensor, mask: torch.Tensor):
        x   = self.transformer(x, src_key_padding_mask=~mask)
        emb = self.pool(x, mask)
        return self.head(emb), emb  # (B, num_traits), (B, in_dim)