Blame for _net.py

"""Codecfake W2V2-AASIST (Xie et al., arXiv 2405.04880) — inference port.

Source: https://github.com/xieyuankun/Codecfake @ 72c32e2840eef95fbeade349425fa9c5391fa11e (NO LICENSE).
  model.py: GraphAttentionLayer, HtrgGraphAttentionLayer, GraphPool, Residual_block, W2VAASIST
  generate_score.py / preprocess.py: pad_dataset (first 64600 samples, repeat-tile), HF
  Wav2Vec2FeatureExtractor (do_normalize) + frozen facebook/wav2vec2-xls-r-300m, hidden_states[5]
  -> unsqueeze(0) -> transpose(2,3) = (B, 1, 1024, T) -> W2VAASIST -> (last_hidden, logits).

Edits vs the source (inference-only, maths unchanged):
  * removed `print(x.shape)` in W2VAASIST.forward and the unused imports (pytorch_model_summary, ...);
  * Residual_block: the source computes bn1->selu for non-first blocks and then applies conv1 to `x`,
    discarding that result; the dead computation is dropped (identical output; bn1 weights kept so the
    checkpoint loads strictly);
  * the frozen XLS-R front-end is truncated to its first 5 transformer layers: hidden_states[5] of the
    24-layer HF model is the output of layer index 4 *before* any final LayerNorm, so `XLSRLayer5`
    runs feature encoder -> feature projection -> pos-conv -> layers[0..4] and returns that tensor
    (bit-identical to the full model's hidden_states[5], verified in implementation-notes.md);
  * the HF feature-extractor normalisation is reproduced per row in torch (zero mean, unit population
    variance, eps 1e-7), so a whole batch is normalised on the GPU;
  * checkpoints are pickled whole modules (torch.save(model)); `load_pickled_state` unpickles them with an
    allow-list unpickler that maps the source's `model.*` classes to a bare nn.Module stub (no source
    code runs) and returns the module's state_dict.
"""
from __future__ import annotations

import pickle
from typing import Union

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F

WINDOW = 64600
XLSR_REPO = "facebook/wav2vec2-xls-r-300m"
XLSR_REV = "1a640f32ac3e39899438a2931f9924c02f080a54"
N_LAYERS = 5  # hidden_states[5]


def pad_repeat(x: np.ndarray, cut: int = WINDOW) -> np.ndarray:
    """Source pad_dataset: first `cut` samples; shorter clips tiled (int(cut/len)+1 repeats) and cropped."""
    n = x.shape[0]
    if n >= cut:
        return x[:cut]
    return np.tile(x, int(cut / n) + 1)[:cut]


def normalize(x: torch.Tensor) -> torch.Tensor:
    """Wav2Vec2FeatureExtractor zero_mean_unit_var_norm, per row: (x - mean) / sqrt(var + 1e-7)."""
    m = x.mean(dim=-1, keepdim=True)
    v = x.var(dim=-1, keepdim=True, unbiased=False)
    return (x - m) / torch.sqrt(v + 1e-7)


# ----------------------------------------------------------------------------- XLS-R front-end
class XLSRLayer5(nn.Module):
    """Frozen XLS-R 300M, first 5 transformer layers; returns HF hidden_states[5] (B, T, 1024)."""

    def __init__(self, w2v):
        super().__init__()
        self.feature_extractor = w2v.feature_extractor
        self.feature_projection = w2v.feature_projection
        self.pos_conv_embed = w2v.encoder.pos_conv_embed
        self.layers = nn.ModuleList(list(w2v.encoder.layers)[:N_LAYERS])

    def forward(self, wav: torch.Tensor) -> torch.Tensor:
        h = self.feature_extractor(wav).transpose(1, 2)
        h, _ = self.feature_projection(h)
        h = h + self.pos_conv_embed(h)
        for layer in self.layers:
            h = layer(h, attention_mask=None)[0]
        return h


def build_xlsr(local_dir: str | None = None) -> XLSRLayer5:
    """facebook/wav2vec2-xls-r-300m @ XLSR_REV, loaded with only N_LAYERS transformer layers (asserted)."""
    from transformers import Wav2Vec2Config, Wav2Vec2Model
    cfg = Wav2Vec2Config.from_pretrained(XLSR_REPO, revision=XLSR_REV)
    assert cfg.do_stable_layer_norm and cfg.num_hidden_layers == 24 and cfg.hidden_size == 1024
    cfg.num_hidden_layers = N_LAYERS
    w2v, info = Wav2Vec2Model.from_pretrained(XLSR_REPO, revision=XLSR_REV, config=cfg,
                                             attn_implementation="eager", output_loading_info=True)
    assert not info["missing_keys"], info["missing_keys"]
    unexp = [k.removeprefix("wav2vec2.") for k in info["unexpected_keys"]]
    bad = [k for k in unexp  # allowed: dropped layers 5..23 + the pre-training-only heads
           if not (k.startswith(("encoder.layers.", "quantizer.", "project_q.", "project_hid."))
                   and not any(k.startswith(f"encoder.layers.{i}.") for i in range(N_LAYERS)))]
    assert sum(k.startswith("encoder.layers.") for k in unexp) == 16 * (24 - N_LAYERS)
    assert not bad, bad
    return XLSRLayer5(w2v)


# ----------------------------------------------------------------------------- AASIST head (model.py)
class GraphAttentionLayer(nn.Module):
    def __init__(self, in_dim, out_dim, **kwargs):
        super().__init__()
        self.att_proj = nn.Linear(in_dim, out_dim)
        self.att_weight = self._init_new_params(out_dim, 1)
        self.proj_with_att = nn.Linear(in_dim, out_dim)
        self.proj_without_att = nn.Linear(in_dim, out_dim)
        self.bn = nn.BatchNorm1d(out_dim)
        self.input_drop = nn.Dropout(p=0.2)
        self.act = nn.SELU(inplace=True)
        self.temp = kwargs.get("temperature", 1.)

    def forward(self, x):
        x = self.input_drop(x)
        att_map = self._derive_att_map(x)
        x = self._project(x, att_map)
        x = self._apply_BN(x)
        return self.act(x)

    def _pairwise_mul_nodes(self, x):
        nb_nodes = x.size(1)
        x = x.unsqueeze(2).expand(-1, -1, nb_nodes, -1)
        return x * x.transpose(1, 2)

    def _derive_att_map(self, x):
        att_map = self._pairwise_mul_nodes(x)
        att_map = torch.tanh(self.att_proj(att_map))
        att_map = torch.matmul(att_map, self.att_weight)
        att_map = att_map / self.temp
        return F.softmax(att_map, dim=-2)

    def _project(self, x, att_map):
        return self.proj_with_att(torch.matmul(att_map.squeeze(-1), x)) + self.proj_without_att(x)

    def _apply_BN(self, x):
        org_size = x.size()
        return self.bn(x.view(-1, org_size[-1])).view(org_size)

    def _init_new_params(self, *size):
        out = nn.Parameter(torch.FloatTensor(*size))
        nn.init.xavier_normal_(out)
        return out


class HtrgGraphAttentionLayer(nn.Module):
    def __init__(self, in_dim, out_dim, **kwargs):
        super().__init__()
        self.proj_type1 = nn.Linear(in_dim, in_dim)
        self.proj_type2 = nn.Linear(in_dim, in_dim)
        self.att_proj = nn.Linear(in_dim, out_dim)
        self.att_projM = nn.Linear(in_dim, out_dim)
        self.att_weight11 = self._init_new_params(out_dim, 1)
        self.att_weight22 = self._init_new_params(out_dim, 1)
        self.att_weight12 = self._init_new_params(out_dim, 1)
        self.att_weightM = self._init_new_params(out_dim, 1)
        self.proj_with_att = nn.Linear(in_dim, out_dim)
        self.proj_without_att = nn.Linear(in_dim, out_dim)
        self.proj_with_attM = nn.Linear(in_dim, out_dim)
        self.proj_without_attM = nn.Linear(in_dim, out_dim)
        self.bn = nn.BatchNorm1d(out_dim)
        self.input_drop = nn.Dropout(p=0.2)
        self.act = nn.SELU(inplace=True)
        self.temp = kwargs.get("temperature", 1.)

    def forward(self, x1, x2, master=None):
        num_type1 = x1.size(1)
        num_type2 = x2.size(1)
        x1 = self.proj_type1(x1)
        x2 = self.proj_type2(x2)
        x = torch.cat([x1, x2], dim=1)
        if master is None:
            master = torch.mean(x, dim=1, keepdim=True)
        x = self.input_drop(x)
        att_map = self._derive_att_map(x, num_type1, num_type2)
        master = self._update_master(x, master)
        x = self._project(x, att_map)
        x = self._apply_BN(x)
        x = self.act(x)
        x1 = x.narrow(1, 0, num_type1)
        x2 = x.narrow(1, num_type1, num_type2)
        return x1, x2, master

    def _update_master(self, x, master):
        att_map = self._derive_att_map_master(x, master)
        return self._project_master(x, master, att_map)

    def _pairwise_mul_nodes(self, x):
        nb_nodes = x.size(1)
        x = x.unsqueeze(2).expand(-1, -1, nb_nodes, -1)
        return x * x.transpose(1, 2)

    def _derive_att_map_master(self, x, master):
        att_map = x * master
        att_map = torch.tanh(self.att_projM(att_map))
        att_map = torch.matmul(att_map, self.att_weightM)
        att_map = att_map / self.temp
        return F.softmax(att_map, dim=-2)

    def _derive_att_map(self, x, num_type1, num_type2):
        att_map = self._pairwise_mul_nodes(x)
        att_map = torch.tanh(self.att_proj(att_map))
        att_board = torch.zeros_like(att_map[:, :, :, 0]).unsqueeze(-1)
        att_board[:, :num_type1, :num_type1, :] = torch.matmul(
            att_map[:, :num_type1, :num_type1, :], self.att_weight11)
        att_board[:, num_type1:, num_type1:, :] = torch.matmul(
            att_map[:, num_type1:, num_type1:, :], self.att_weight22)
        att_board[:, :num_type1, num_type1:, :] = torch.matmul(
            att_map[:, :num_type1, num_type1:, :], self.att_weight12)
        att_board[:, num_type1:, :num_type1, :] = torch.matmul(
            att_map[:, num_type1:, :num_type1, :], self.att_weight12)
        att_map = att_board / self.temp
        return F.softmax(att_map, dim=-2)

    def _project(self, x, att_map):
        return self.proj_with_att(torch.matmul(att_map.squeeze(-1), x)) + self.proj_without_att(x)

    def _project_master(self, x, master, att_map):
        x1 = self.proj_with_attM(torch.matmul(att_map.squeeze(-1).unsqueeze(1), x))
        return x1 + self.proj_without_attM(master)

    def _apply_BN(self, x):
        org_size = x.size()
        return self.bn(x.view(-1, org_size[-1])).view(org_size)

    def _init_new_params(self, *size):
        out = nn.Parameter(torch.FloatTensor(*size))
        nn.init.xavier_normal_(out)
        return out


class GraphPool(nn.Module):
    def __init__(self, k: float, in_dim: int, p: Union[float, int]):
        super().__init__()
        self.k = k
        self.sigmoid = nn.Sigmoid()
        self.proj = nn.Linear(in_dim, 1)
        self.drop = nn.Dropout(p=p) if p > 0 else nn.Identity()
        self.in_dim = in_dim

    def forward(self, h):
        scores = self.sigmoid(self.proj(self.drop(h)))
        return self.top_k_graph(scores, h, self.k)

    def top_k_graph(self, scores, h, k):
        _, n_nodes, n_feat = h.size()
        n_nodes = max(int(n_nodes * k), 1)
        _, idx = torch.topk(scores, n_nodes, dim=1)
        idx = idx.expand(-1, -1, n_feat)
        h = h * scores
        return torch.gather(h, 1, idx)


class Residual_block(nn.Module):
    def __init__(self, nb_filts, first=False):
        super().__init__()
        self.first = first
        if not self.first:
            self.bn1 = nn.BatchNorm2d(num_features=nb_filts[0])
        self.conv1 = nn.Conv2d(nb_filts[0], nb_filts[1], kernel_size=(2, 3), padding=(1, 1), stride=1)
        self.selu = nn.SELU(inplace=True)
        self.bn2 = nn.BatchNorm2d(num_features=nb_filts[1])
        self.conv2 = nn.Conv2d(nb_filts[1], nb_filts[1], kernel_size=(2, 3), padding=(0, 1), stride=1)
        if nb_filts[0] != nb_filts[1]:
            self.downsample = True
            self.conv_downsample = nn.Conv2d(nb_filts[0], nb_filts[1], padding=(0, 1), kernel_size=(1, 3), stride=1)
        else:
            self.downsample = False

    def forward(self, x):
        identity = x
        # source: bn1/selu computed for non-first blocks, but conv1 is applied to x (kept verbatim; the
        # inplace SELU acts on bn1's fresh output, so x is untouched)
        out = self.conv1(x)
        out = self.bn2(out)
        out = self.selu(out)
        out = self.conv2(out)
        if self.downsample:
            identity = self.conv_downsample(identity)
        out += identity
        return out


class W2VAASIST(nn.Module):
    def __init__(self):
        super().__init__()
        filts = [128, [1, 32], [32, 32], [32, 64], [64, 64]]
        gat_dims = [64, 32]
        pool_ratios = [0.5, 0.5, 0.5, 0.5]
        temperatures = [2.0, 2.0, 100.0, 100.0]
        self.first_bn = nn.BatchNorm2d(num_features=1)
        self.first_bn1 = nn.BatchNorm2d(num_features=64)
        self.drop = nn.Dropout(0.5, inplace=True)
        self.drop_way = nn.Dropout(0.2, inplace=True)
        self.selu = nn.SELU(inplace=True)
        self.encoder = nn.Sequential(
            nn.Sequential(Residual_block(nb_filts=filts[1], first=True)),
            nn.Sequential(Residual_block(nb_filts=filts[2])),
            nn.Sequential(Residual_block(nb_filts=filts[3])),
            nn.Sequential(Residual_block(nb_filts=filts[4])),
            nn.Sequential(Residual_block(nb_filts=filts[4])),
            nn.Sequential(Residual_block(nb_filts=filts[4])))
        self.LL = nn.Linear(1024, 128)
        self.attention = nn.Sequential(
            nn.Conv2d(64, 128, kernel_size=(1, 1)),
            nn.SELU(inplace=True),
            nn.BatchNorm2d(128),
            nn.Conv2d(128, 64, kernel_size=(1, 1)),
        )
        self.pos_S = nn.Parameter(torch.randn(1, 42, filts[-1][-1]))
        self.master1 = nn.Parameter(torch.randn(1, 1, gat_dims[0]))
        self.master2 = nn.Parameter(torch.randn(1, 1, gat_dims[0]))
        self.GAT_layer_S = GraphAttentionLayer(filts[-1][-1], gat_dims[0], temperature=temperatures[0])
        self.GAT_layer_T = GraphAttentionLayer(filts[-1][-1], gat_dims[0], temperature=temperatures[1])
        self.HtrgGAT_layer_ST11 = HtrgGraphAttentionLayer(gat_dims[0], gat_dims[1], temperature=temperatures[2])
        self.HtrgGAT_layer_ST12 = HtrgGraphAttentionLayer(gat_dims[1], gat_dims[1], temperature=temperatures[2])
        self.HtrgGAT_layer_ST21 = HtrgGraphAttentionLayer(gat_dims[0], gat_dims[1], temperature=temperatures[2])
        self.HtrgGAT_layer_ST22 = HtrgGraphAttentionLayer(gat_dims[1], gat_dims[1], temperature=temperatures[2])
        self.pool_S = GraphPool(pool_ratios[0], gat_dims[0], 0.3)
        self.pool_T = GraphPool(pool_ratios[1], gat_dims[0], 0.3)
        self.pool_hS1 = GraphPool(pool_ratios[2], gat_dims[1], 0.3)
        self.pool_hT1 = GraphPool(pool_ratios[2], gat_dims[1], 0.3)
        self.pool_hS2 = GraphPool(pool_ratios[2], gat_dims[1], 0.3)
        self.pool_hT2 = GraphPool(pool_ratios[2], gat_dims[1], 0.3)
        self.out_layer = nn.Linear(5 * gat_dims[1], 2)

    def forward(self, x):
        x = x.squeeze(dim=1)
        x = x.transpose(1, 2)
        x = self.LL(x)
        x = x.transpose(1, 2)
        x = x.unsqueeze(dim=1)
        x = F.max_pool2d(x, (3, 3))
        x = self.first_bn(x)
        x = self.selu(x)
        x = self.encoder(x)
        x = self.first_bn1(x)
        x = self.selu(x)
        w = self.attention(x)
        w1 = F.softmax(w, dim=-1)
        m = torch.sum(x * w1, dim=-1)
        e_S = m.transpose(1, 2) + self.pos_S
        gat_S = self.GAT_layer_S(e_S)
        out_S = self.pool_S(gat_S)
        w2 = F.softmax(w, dim=-2)
        m1 = torch.sum(x * w2, dim=-2)
        e_T = m1.transpose(1, 2)
        gat_T = self.GAT_layer_T(e_T)
        out_T = self.pool_T(gat_T)
        master1 = self.master1.expand(x.size(0), -1, -1)
        master2 = self.master2.expand(x.size(0), -1, -1)
        out_T1, out_S1, master1 = self.HtrgGAT_layer_ST11(out_T, out_S, master=self.master1)
        out_S1 = self.pool_hS1(out_S1)
        out_T1 = self.pool_hT1(out_T1)
        out_T_aug, out_S_aug, master_aug = self.HtrgGAT_layer_ST12(out_T1, out_S1, master=master1)
        out_T1 = out_T1 + out_T_aug
        out_S1 = out_S1 + out_S_aug
        master1 = master1 + master_aug
        out_T2, out_S2, master2 = self.HtrgGAT_layer_ST21(out_T, out_S, master=self.master2)
        out_S2 = self.pool_hS2(out_S2)
        out_T2 = self.pool_hT2(out_T2)
        out_T_aug, out_S_aug, master_aug = self.HtrgGAT_layer_ST22(out_T2, out_S2, master=master2)
        out_T2 = out_T2 + out_T_aug
        out_S2 = out_S2 + out_S_aug
        master2 = master2 + master_aug
        out_T1 = self.drop_way(out_T1)
        out_T2 = self.drop_way(out_T2)
        out_S1 = self.drop_way(out_S1)
        out_S2 = self.drop_way(out_S2)
        master1 = self.drop_way(master1)
        master2 = self.drop_way(master2)
        out_T = torch.max(out_T1, out_T2)
        out_S = torch.max(out_S1, out_S2)
        master = torch.max(master1, master2)
        T_max, _ = torch.max(torch.abs(out_T), dim=1)
        T_avg = torch.mean(out_T, dim=1)
        S_max, _ = torch.max(torch.abs(out_S), dim=1)
        S_avg = torch.mean(out_S, dim=1)
        last_hidden = torch.cat([T_max, T_avg, S_max, S_avg, master.squeeze(1)], dim=1)
        last_hidden = self.drop(last_hidden)
        output = self.out_layer(last_hidden)
        return last_hidden, output


class CodecfakeNet(nn.Module):
    """wav (B, 64600) -> per-row normalisation -> XLS-R hidden_states[5] -> W2VAASIST logits (B, 2)."""

    def __init__(self, xlsr: XLSRLayer5, head: W2VAASIST):
        super().__init__()
        self.xlsr = xlsr
        self.head = head

    def forward(self, wav: torch.Tensor) -> torch.Tensor:
        h = self.xlsr(normalize(wav))                    # (B, T, 1024)
        h = h.unsqueeze(1).transpose(2, 3)               # (B, 1, 1024, T), as generate_score.py
        return self.head(h)[1]


# ----------------------------------------------------------------------------- checkpoint conversion
class _Stub(nn.Module):
    """Stand-in for the source's model.* classes while unpickling (state only, no source code)."""

    def forward(self, *a, **k):  # pragma: no cover
        raise RuntimeError("stub")


_ALLOWED = {
    ("collections", "OrderedDict"), ("builtins", "set"), ("__builtin__", "set"),
    ("torch", "FloatStorage"), ("torch", "LongStorage"),
    ("torch._utils", "_rebuild_parameter"), ("torch._utils", "_rebuild_tensor_v2"),
}
_SOURCE_CLASSES = {"W2VAASIST", "GraphAttentionLayer", "HtrgGraphAttentionLayer", "GraphPool", "Residual_block"}


class _AllowListUnpickler(pickle.Unpickler):
    def find_class(self, module, name):
        if module == "model" and name in _SOURCE_CLASSES:
            return _Stub
        if module.startswith("torch.nn.modules.") or (module, name) in _ALLOWED:
            return super().find_class(module, name)
        raise pickle.UnpicklingError(f"blocked global {module}.{name}")


class _PickleModule:
    Unpickler = _AllowListUnpickler
    load = staticmethod(pickle.load)
    __name__ = "codecfake_allowlist_pickle"


def load_pickled_state(path: str) -> dict:
    """state_dict of the source's torch.save(W2VAASIST) file, via the allow-list unpickler."""
    mod = torch.load(path, map_location="cpu", weights_only=False, pickle_module=_PickleModule)
    assert isinstance(mod, _Stub), type(mod)
    return {k: v.detach().clone().contiguous() for k, v in mod.state_dict().items()}


# ----------------------------------------------------------------------------- weight acquisition
SRC_REPO = "xieyuankun/Codecfake"
SRC_REV = "72c32e2840eef95fbeade349425fa9c5391fa11e"
CKPT_SHA256 = {
    "cotrain_w2v2aasist_CSAM": "86876e2c0eb98d6e60ecf2f1139112206adf47b1562aee72530d0c22e04c644a",
    "codec_w2v2aasist": "5a606d1fbd578dbd214471cead283c236f4204c54d602e945924b3594f8c25a3",
}


def ckpt_url(sub: str) -> str:
    return f"https://raw.githubusercontent.com/{SRC_REPO}/{SRC_REV}/pretrained_model/{sub}/anti-spoofing_feat_model.pt"


def head_state(sub: str, model_dir: str) -> dict:
    """W2VAASIST state_dict for pretrained_model/<sub>. No upstream license -> not re-hosted: on first use the
    authors' pickled module is downloaded from the commit-pinned GitHub raw URL (sha256-checked) into
    ../_ckpt_dl/xieyuankun__Codecfake/<rev>/, converted once with the allow-list unpickler, and cached as
    <model_dir>/head_<sub>.safetensors (loaded on every later run)."""
    import hashlib
    import os
    import urllib.request
    from safetensors.torch import load_file, save_file
    cache = os.path.join(model_dir, f"head_{sub}.safetensors")
    if not os.path.exists(cache):
        pt = os.path.join(model_dir, "..", "_ckpt_dl", "xieyuankun__Codecfake", SRC_REV, "pretrained_model",
                          sub, "anti-spoofing_feat_model.pt")
        if not os.path.exists(pt):
            os.makedirs(os.path.dirname(pt), exist_ok=True)
            urllib.request.urlretrieve(ckpt_url(sub), pt + ".part")
            os.replace(pt + ".part", pt)
        digest = hashlib.sha256(open(pt, "rb").read()).hexdigest()
        if digest != CKPT_SHA256[sub]:
            raise RuntimeError(f"{pt}: sha256 {digest} != pinned {CKPT_SHA256[sub]}")
        save_file(load_pickled_state(pt), cache + ".part", metadata={"source": ckpt_url(sub), "sha256": digest})
        os.replace(cache + ".part", cache)
    return load_file(cache)